1 /*
2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/classLoaderData.hpp"
27 #include "classfile/stringTable.hpp"
28 #include "classfile/symbolTable.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "code/codeCache.hpp"
31 #include "gc/cms/cmsCollectorPolicy.hpp"
32 #include "gc/cms/cmsHeap.hpp"
33 #include "gc/cms/cmsOopClosures.inline.hpp"
34 #include "gc/cms/compactibleFreeListSpace.hpp"
35 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp"
36 #include "gc/cms/concurrentMarkSweepThread.hpp"
37 #include "gc/cms/parNewGeneration.hpp"
38 #include "gc/cms/vmCMSOperations.hpp"
39 #include "gc/serial/genMarkSweep.hpp"
40 #include "gc/serial/tenuredGeneration.hpp"
41 #include "gc/shared/adaptiveSizePolicy.hpp"
42 #include "gc/shared/cardGeneration.inline.hpp"
43 #include "gc/shared/cardTableRS.hpp"
44 #include "gc/shared/collectedHeap.inline.hpp"
45 #include "gc/shared/collectorCounters.hpp"
46 #include "gc/shared/collectorPolicy.hpp"
47 #include "gc/shared/gcLocker.hpp"
48 #include "gc/shared/gcPolicyCounters.hpp"
49 #include "gc/shared/gcTimer.hpp"
50 #include "gc/shared/gcTrace.hpp"
51 #include "gc/shared/gcTraceTime.inline.hpp"
52 #include "gc/shared/genCollectedHeap.hpp"
53 #include "gc/shared/genOopClosures.inline.hpp"
54 #include "gc/shared/isGCActiveMark.hpp"
55 #include "gc/shared/referencePolicy.hpp"
56 #include "gc/shared/strongRootsScope.hpp"
57 #include "gc/shared/taskqueue.inline.hpp"
58 #include "gc/shared/weakProcessor.hpp"
59 #include "logging/log.hpp"
60 #include "logging/logStream.hpp"
61 #include "memory/allocation.hpp"
62 #include "memory/iterator.inline.hpp"
63 #include "memory/padded.hpp"
64 #include "memory/resourceArea.hpp"
65 #include "oops/access.inline.hpp"
66 #include "oops/oop.inline.hpp"
67 #include "prims/jvmtiExport.hpp"
68 #include "runtime/atomic.hpp"
69 #include "runtime/globals_extension.hpp"
70 #include "runtime/handles.inline.hpp"
71 #include "runtime/java.hpp"
72 #include "runtime/orderAccess.inline.hpp"
73 #include "runtime/timer.hpp"
74 #include "runtime/vmThread.hpp"
75 #include "services/memoryService.hpp"
76 #include "services/runtimeService.hpp"
77 #include "utilities/align.hpp"
78 #include "utilities/stack.inline.hpp"
79
80 // statics
81 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL;
82 bool CMSCollector::_full_gc_requested = false;
83 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc;
84
85 //////////////////////////////////////////////////////////////////
86 // In support of CMS/VM thread synchronization
87 //////////////////////////////////////////////////////////////////
88 // We split use of the CGC_lock into 2 "levels".
89 // The low-level locking is of the usual CGC_lock monitor. We introduce
90 // a higher level "token" (hereafter "CMS token") built on top of the
91 // low level monitor (hereafter "CGC lock").
92 // The token-passing protocol gives priority to the VM thread. The
93 // CMS-lock doesn't provide any fairness guarantees, but clients
94 // should ensure that it is only held for very short, bounded
95 // durations.
96 //
97 // When either of the CMS thread or the VM thread is involved in
98 // collection operations during which it does not want the other
99 // thread to interfere, it obtains the CMS token.
100 //
101 // If either thread tries to get the token while the other has
102 // it, that thread waits. However, if the VM thread and CMS thread
103 // both want the token, then the VM thread gets priority while the
104 // CMS thread waits. This ensures, for instance, that the "concurrent"
105 // phases of the CMS thread's work do not block out the VM thread
106 // for long periods of time as the CMS thread continues to hog
107 // the token. (See bug 4616232).
108 //
109 // The baton-passing functions are, however, controlled by the
110 // flags _foregroundGCShouldWait and _foregroundGCIsActive,
111 // and here the low-level CMS lock, not the high level token,
112 // ensures mutual exclusion.
113 //
114 // Two important conditions that we have to satisfy:
115 // 1. if a thread does a low-level wait on the CMS lock, then it
116 // relinquishes the CMS token if it were holding that token
117 // when it acquired the low-level CMS lock.
118 // 2. any low-level notifications on the low-level lock
119 // should only be sent when a thread has relinquished the token.
120 //
121 // In the absence of either property, we'd have potential deadlock.
122 //
123 // We protect each of the CMS (concurrent and sequential) phases
124 // with the CMS _token_, not the CMS _lock_.
125 //
126 // The only code protected by CMS lock is the token acquisition code
127 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the
128 // baton-passing code.
129 //
130 // Unfortunately, i couldn't come up with a good abstraction to factor and
131 // hide the naked CGC_lock manipulation in the baton-passing code
132 // further below. That's something we should try to do. Also, the proof
133 // of correctness of this 2-level locking scheme is far from obvious,
134 // and potentially quite slippery. We have an uneasy suspicion, for instance,
135 // that there may be a theoretical possibility of delay/starvation in the
136 // low-level lock/wait/notify scheme used for the baton-passing because of
137 // potential interference with the priority scheme embodied in the
138 // CMS-token-passing protocol. See related comments at a CGC_lock->wait()
139 // invocation further below and marked with "XXX 20011219YSR".
140 // Indeed, as we note elsewhere, this may become yet more slippery
141 // in the presence of multiple CMS and/or multiple VM threads. XXX
142
143 class CMSTokenSync: public StackObj {
144 private:
145 bool _is_cms_thread;
146 public:
147 CMSTokenSync(bool is_cms_thread):
148 _is_cms_thread(is_cms_thread) {
149 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(),
150 "Incorrect argument to constructor");
151 ConcurrentMarkSweepThread::synchronize(_is_cms_thread);
152 }
153
154 ~CMSTokenSync() {
155 assert(_is_cms_thread ?
156 ConcurrentMarkSweepThread::cms_thread_has_cms_token() :
157 ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
158 "Incorrect state");
159 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread);
160 }
161 };
162
163 // Convenience class that does a CMSTokenSync, and then acquires
164 // upto three locks.
165 class CMSTokenSyncWithLocks: public CMSTokenSync {
166 private:
167 // Note: locks are acquired in textual declaration order
168 // and released in the opposite order
169 MutexLockerEx _locker1, _locker2, _locker3;
170 public:
171 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1,
172 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL):
173 CMSTokenSync(is_cms_thread),
174 _locker1(mutex1, Mutex::_no_safepoint_check_flag),
175 _locker2(mutex2, Mutex::_no_safepoint_check_flag),
176 _locker3(mutex3, Mutex::_no_safepoint_check_flag)
177 { }
178 };
179
180
181 //////////////////////////////////////////////////////////////////
182 // Concurrent Mark-Sweep Generation /////////////////////////////
183 //////////////////////////////////////////////////////////////////
184
185 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;)
186
187 // This struct contains per-thread things necessary to support parallel
188 // young-gen collection.
189 class CMSParGCThreadState: public CHeapObj<mtGC> {
190 public:
191 CompactibleFreeListSpaceLAB lab;
192 PromotionInfo promo;
193
194 // Constructor.
195 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) {
196 promo.setSpace(cfls);
197 }
198 };
199
200 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration(
201 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) :
202 CardGeneration(rs, initial_byte_size, ct),
203 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))),
204 _did_compact(false)
205 {
206 HeapWord* bottom = (HeapWord*) _virtual_space.low();
207 HeapWord* end = (HeapWord*) _virtual_space.high();
208
209 _direct_allocated_words = 0;
210 NOT_PRODUCT(
211 _numObjectsPromoted = 0;
212 _numWordsPromoted = 0;
213 _numObjectsAllocated = 0;
214 _numWordsAllocated = 0;
215 )
216
217 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end));
218 NOT_PRODUCT(debug_cms_space = _cmsSpace;)
219 _cmsSpace->_old_gen = this;
220
221 _gc_stats = new CMSGCStats();
222
223 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass
224 // offsets match. The ability to tell free chunks from objects
225 // depends on this property.
226 debug_only(
227 FreeChunk* junk = NULL;
228 assert(UseCompressedClassPointers ||
229 junk->prev_addr() == (void*)(oop(junk)->klass_addr()),
230 "Offset of FreeChunk::_prev within FreeChunk must match"
231 " that of OopDesc::_klass within OopDesc");
232 )
233
234 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC);
235 for (uint i = 0; i < ParallelGCThreads; i++) {
236 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace());
237 }
238
239 _incremental_collection_failed = false;
240 // The "dilatation_factor" is the expansion that can occur on
241 // account of the fact that the minimum object size in the CMS
242 // generation may be larger than that in, say, a contiguous young
243 // generation.
244 // Ideally, in the calculation below, we'd compute the dilatation
245 // factor as: MinChunkSize/(promoting_gen's min object size)
246 // Since we do not have such a general query interface for the
247 // promoting generation, we'll instead just use the minimum
248 // object size (which today is a header's worth of space);
249 // note that all arithmetic is in units of HeapWords.
250 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking");
251 assert(_dilatation_factor >= 1.0, "from previous assert");
252 }
253
254
255 // The field "_initiating_occupancy" represents the occupancy percentage
256 // at which we trigger a new collection cycle. Unless explicitly specified
257 // via CMSInitiatingOccupancyFraction (argument "io" below), it
258 // is calculated by:
259 //
260 // Let "f" be MinHeapFreeRatio in
261 //
262 // _initiating_occupancy = 100-f +
263 // f * (CMSTriggerRatio/100)
264 // where CMSTriggerRatio is the argument "tr" below.
265 //
266 // That is, if we assume the heap is at its desired maximum occupancy at the
267 // end of a collection, we let CMSTriggerRatio of the (purported) free
268 // space be allocated before initiating a new collection cycle.
269 //
270 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) {
271 assert(io <= 100 && tr <= 100, "Check the arguments");
272 if (io >= 0) {
273 _initiating_occupancy = (double)io / 100.0;
274 } else {
275 _initiating_occupancy = ((100 - MinHeapFreeRatio) +
276 (double)(tr * MinHeapFreeRatio) / 100.0)
277 / 100.0;
278 }
279 }
280
281 void ConcurrentMarkSweepGeneration::ref_processor_init() {
282 assert(collector() != NULL, "no collector");
283 collector()->ref_processor_init();
284 }
285
286 void CMSCollector::ref_processor_init() {
287 if (_ref_processor == NULL) {
288 // Allocate and initialize a reference processor
289 _ref_processor =
290 new ReferenceProcessor(_span, // span
291 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing
292 ParallelGCThreads, // mt processing degree
293 _cmsGen->refs_discovery_is_mt(), // mt discovery
294 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree
295 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic
296 &_is_alive_closure); // closure for liveness info
297 // Initialize the _ref_processor field of CMSGen
298 _cmsGen->set_ref_processor(_ref_processor);
299
300 }
301 }
302
303 AdaptiveSizePolicy* CMSCollector::size_policy() {
304 return CMSHeap::heap()->size_policy();
305 }
306
307 void ConcurrentMarkSweepGeneration::initialize_performance_counters() {
308
309 const char* gen_name = "old";
310 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy();
311 // Generation Counters - generation 1, 1 subspace
312 _gen_counters = new GenerationCounters(gen_name, 1, 1,
313 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space);
314
315 _space_counters = new GSpaceCounters(gen_name, 0,
316 _virtual_space.reserved_size(),
317 this, _gen_counters);
318 }
319
320 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha):
321 _cms_gen(cms_gen)
322 {
323 assert(alpha <= 100, "bad value");
324 _saved_alpha = alpha;
325
326 // Initialize the alphas to the bootstrap value of 100.
327 _gc0_alpha = _cms_alpha = 100;
328
329 _cms_begin_time.update();
330 _cms_end_time.update();
331
332 _gc0_duration = 0.0;
333 _gc0_period = 0.0;
334 _gc0_promoted = 0;
335
336 _cms_duration = 0.0;
337 _cms_period = 0.0;
338 _cms_allocated = 0;
339
340 _cms_used_at_gc0_begin = 0;
341 _cms_used_at_gc0_end = 0;
342 _allow_duty_cycle_reduction = false;
343 _valid_bits = 0;
344 }
345
346 double CMSStats::cms_free_adjustment_factor(size_t free) const {
347 // TBD: CR 6909490
348 return 1.0;
349 }
350
351 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) {
352 }
353
354 // If promotion failure handling is on use
355 // the padded average size of the promotion for each
356 // young generation collection.
357 double CMSStats::time_until_cms_gen_full() const {
358 size_t cms_free = _cms_gen->cmsSpace()->free();
359 CMSHeap* heap = CMSHeap::heap();
360 size_t expected_promotion = MIN2(heap->young_gen()->capacity(),
361 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average());
362 if (cms_free > expected_promotion) {
363 // Start a cms collection if there isn't enough space to promote
364 // for the next young collection. Use the padded average as
365 // a safety factor.
366 cms_free -= expected_promotion;
367
368 // Adjust by the safety factor.
369 double cms_free_dbl = (double)cms_free;
370 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0;
371 // Apply a further correction factor which tries to adjust
372 // for recent occurance of concurrent mode failures.
373 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free);
374 cms_free_dbl = cms_free_dbl * cms_adjustment;
375
376 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT,
377 cms_free, expected_promotion);
378 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0);
379 // Add 1 in case the consumption rate goes to zero.
380 return cms_free_dbl / (cms_consumption_rate() + 1.0);
381 }
382 return 0.0;
383 }
384
385 // Compare the duration of the cms collection to the
386 // time remaining before the cms generation is empty.
387 // Note that the time from the start of the cms collection
388 // to the start of the cms sweep (less than the total
389 // duration of the cms collection) can be used. This
390 // has been tried and some applications experienced
391 // promotion failures early in execution. This was
392 // possibly because the averages were not accurate
393 // enough at the beginning.
394 double CMSStats::time_until_cms_start() const {
395 // We add "gc0_period" to the "work" calculation
396 // below because this query is done (mostly) at the
397 // end of a scavenge, so we need to conservatively
398 // account for that much possible delay
399 // in the query so as to avoid concurrent mode failures
400 // due to starting the collection just a wee bit too
401 // late.
402 double work = cms_duration() + gc0_period();
403 double deadline = time_until_cms_gen_full();
404 // If a concurrent mode failure occurred recently, we want to be
405 // more conservative and halve our expected time_until_cms_gen_full()
406 if (work > deadline) {
407 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ",
408 cms_duration(), gc0_period(), time_until_cms_gen_full());
409 return 0.0;
410 }
411 return work - deadline;
412 }
413
414 #ifndef PRODUCT
415 void CMSStats::print_on(outputStream *st) const {
416 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha);
417 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT,
418 gc0_duration(), gc0_period(), gc0_promoted());
419 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT,
420 cms_duration(), cms_period(), cms_allocated());
421 st->print(",cms_since_beg=%g,cms_since_end=%g",
422 cms_time_since_begin(), cms_time_since_end());
423 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT,
424 _cms_used_at_gc0_begin, _cms_used_at_gc0_end);
425
426 if (valid()) {
427 st->print(",promo_rate=%g,cms_alloc_rate=%g",
428 promotion_rate(), cms_allocation_rate());
429 st->print(",cms_consumption_rate=%g,time_until_full=%g",
430 cms_consumption_rate(), time_until_cms_gen_full());
431 }
432 st->cr();
433 }
434 #endif // #ifndef PRODUCT
435
436 CMSCollector::CollectorState CMSCollector::_collectorState =
437 CMSCollector::Idling;
438 bool CMSCollector::_foregroundGCIsActive = false;
439 bool CMSCollector::_foregroundGCShouldWait = false;
440
441 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
442 CardTableRS* ct,
443 ConcurrentMarkSweepPolicy* cp):
444 _cmsGen(cmsGen),
445 _ct(ct),
446 _ref_processor(NULL), // will be set later
447 _conc_workers(NULL), // may be set later
448 _abort_preclean(false),
449 _start_sampling(false),
450 _between_prologue_and_epilogue(false),
451 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"),
452 _modUnionTable((CardTable::card_shift - LogHeapWordSize),
453 -1 /* lock-free */, "No_lock" /* dummy */),
454 _modUnionClosurePar(&_modUnionTable),
455 // Adjust my span to cover old (cms) gen
456 _span(cmsGen->reserved()),
457 // Construct the is_alive_closure with _span & markBitMap
458 _is_alive_closure(_span, &_markBitMap),
459 _restart_addr(NULL),
460 _overflow_list(NULL),
461 _stats(cmsGen),
462 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true,
463 //verify that this lock should be acquired with safepoint check.
464 Monitor::_safepoint_check_sometimes)),
465 _eden_chunk_array(NULL), // may be set in ctor body
466 _eden_chunk_capacity(0), // -- ditto --
467 _eden_chunk_index(0), // -- ditto --
468 _survivor_plab_array(NULL), // -- ditto --
469 _survivor_chunk_array(NULL), // -- ditto --
470 _survivor_chunk_capacity(0), // -- ditto --
471 _survivor_chunk_index(0), // -- ditto --
472 _ser_pmc_preclean_ovflw(0),
473 _ser_kac_preclean_ovflw(0),
474 _ser_pmc_remark_ovflw(0),
475 _par_pmc_remark_ovflw(0),
476 _ser_kac_ovflw(0),
477 _par_kac_ovflw(0),
478 #ifndef PRODUCT
479 _num_par_pushes(0),
480 #endif
481 _collection_count_start(0),
482 _verifying(false),
483 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"),
484 _completed_initialization(false),
485 _collector_policy(cp),
486 _should_unload_classes(CMSClassUnloadingEnabled),
487 _concurrent_cycles_since_last_unload(0),
488 _roots_scanning_options(GenCollectedHeap::SO_None),
489 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
490 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding),
491 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()),
492 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
493 _cms_start_registered(false)
494 {
495 // Now expand the span and allocate the collection support structures
496 // (MUT, marking bit map etc.) to cover both generations subject to
497 // collection.
498
499 // For use by dirty card to oop closures.
500 _cmsGen->cmsSpace()->set_collector(this);
501
502 // Allocate MUT and marking bit map
503 {
504 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag);
505 if (!_markBitMap.allocate(_span)) {
506 log_warning(gc)("Failed to allocate CMS Bit Map");
507 return;
508 }
509 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?");
510 }
511 {
512 _modUnionTable.allocate(_span);
513 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?");
514 }
515
516 if (!_markStack.allocate(MarkStackSize)) {
517 log_warning(gc)("Failed to allocate CMS Marking Stack");
518 return;
519 }
520
521 // Support for multi-threaded concurrent phases
522 if (CMSConcurrentMTEnabled) {
523 if (FLAG_IS_DEFAULT(ConcGCThreads)) {
524 // just for now
525 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4);
526 }
527 if (ConcGCThreads > 1) {
528 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread",
529 ConcGCThreads, true);
530 if (_conc_workers == NULL) {
531 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled");
532 CMSConcurrentMTEnabled = false;
533 } else {
534 _conc_workers->initialize_workers();
535 }
536 } else {
537 CMSConcurrentMTEnabled = false;
538 }
539 }
540 if (!CMSConcurrentMTEnabled) {
541 ConcGCThreads = 0;
542 } else {
543 // Turn off CMSCleanOnEnter optimization temporarily for
544 // the MT case where it's not fixed yet; see 6178663.
545 CMSCleanOnEnter = false;
546 }
547 assert((_conc_workers != NULL) == (ConcGCThreads > 1),
548 "Inconsistency");
549 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
550 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
551
552 // Parallel task queues; these are shared for the
553 // concurrent and stop-world phases of CMS, but
554 // are not shared with parallel scavenge (ParNew).
555 {
556 uint i;
557 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads);
558
559 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled
560 || ParallelRefProcEnabled)
561 && num_queues > 0) {
562 _task_queues = new OopTaskQueueSet(num_queues);
563 if (_task_queues == NULL) {
564 log_warning(gc)("task_queues allocation failure.");
565 return;
566 }
567 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
568 typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
569 for (i = 0; i < num_queues; i++) {
570 PaddedOopTaskQueue *q = new PaddedOopTaskQueue();
571 if (q == NULL) {
572 log_warning(gc)("work_queue allocation failure.");
573 return;
574 }
575 _task_queues->register_queue(i, q);
576 }
577 for (i = 0; i < num_queues; i++) {
578 _task_queues->queue(i)->initialize();
579 _hash_seed[i] = 17; // copied from ParNew
580 }
581 }
582 }
583
584 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio);
585
586 // Clip CMSBootstrapOccupancy between 0 and 100.
587 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0;
588
589 // Now tell CMS generations the identity of their collector
590 ConcurrentMarkSweepGeneration::set_collector(this);
591
592 // Create & start a CMS thread for this CMS collector
593 _cmsThread = ConcurrentMarkSweepThread::start(this);
594 assert(cmsThread() != NULL, "CMS Thread should have been created");
595 assert(cmsThread()->collector() == this,
596 "CMS Thread should refer to this gen");
597 assert(CGC_lock != NULL, "Where's the CGC_lock?");
598
599 // Support for parallelizing young gen rescan
600 CMSHeap* heap = CMSHeap::heap();
601 assert(heap->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew");
602 _young_gen = (ParNewGeneration*)heap->young_gen();
603 if (heap->supports_inline_contig_alloc()) {
604 _top_addr = heap->top_addr();
605 _end_addr = heap->end_addr();
606 assert(_young_gen != NULL, "no _young_gen");
607 _eden_chunk_index = 0;
608 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain;
609 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC);
610 }
611
612 // Support for parallelizing survivor space rescan
613 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) {
614 const size_t max_plab_samples =
615 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize);
616
617 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC);
618 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
619 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC);
620 _survivor_chunk_capacity = max_plab_samples;
621 for (uint i = 0; i < ParallelGCThreads; i++) {
622 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC);
623 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples);
624 assert(cur->end() == 0, "Should be 0");
625 assert(cur->array() == vec, "Should be vec");
626 assert(cur->capacity() == max_plab_samples, "Error");
627 }
628 }
629
630 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;)
631 _gc_counters = new CollectorCounters("CMS", 1);
632 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2);
633 _completed_initialization = true;
634 _inter_sweep_timer.start(); // start of time
635 }
636
637 const char* ConcurrentMarkSweepGeneration::name() const {
638 return "concurrent mark-sweep generation";
639 }
640 void ConcurrentMarkSweepGeneration::update_counters() {
641 if (UsePerfData) {
642 _space_counters->update_all();
643 _gen_counters->update_all();
644 }
645 }
646
647 // this is an optimized version of update_counters(). it takes the
648 // used value as a parameter rather than computing it.
649 //
650 void ConcurrentMarkSweepGeneration::update_counters(size_t used) {
651 if (UsePerfData) {
652 _space_counters->update_used(used);
653 _space_counters->update_capacity();
654 _gen_counters->update_all();
655 }
656 }
657
658 void ConcurrentMarkSweepGeneration::print() const {
659 Generation::print();
660 cmsSpace()->print();
661 }
662
663 #ifndef PRODUCT
664 void ConcurrentMarkSweepGeneration::print_statistics() {
665 cmsSpace()->printFLCensus(0);
666 }
667 #endif
668
669 size_t
670 ConcurrentMarkSweepGeneration::contiguous_available() const {
671 // dld proposes an improvement in precision here. If the committed
672 // part of the space ends in a free block we should add that to
673 // uncommitted size in the calculation below. Will make this
674 // change later, staying with the approximation below for the
675 // time being. -- ysr.
676 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc());
677 }
678
679 size_t
680 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const {
681 return _cmsSpace->max_alloc_in_words() * HeapWordSize;
682 }
683
684 size_t ConcurrentMarkSweepGeneration::max_available() const {
685 return free() + _virtual_space.uncommitted_size();
686 }
687
688 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const {
689 size_t available = max_available();
690 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average();
691 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes);
692 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")",
693 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes);
694 return res;
695 }
696
697 // At a promotion failure dump information on block layout in heap
698 // (cms old generation).
699 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() {
700 Log(gc, promotion) log;
701 if (log.is_trace()) {
702 LogStream ls(log.trace());
703 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls);
704 }
705 }
706
707 void ConcurrentMarkSweepGeneration::reset_after_compaction() {
708 // Clear the promotion information. These pointers can be adjusted
709 // along with all the other pointers into the heap but
710 // compaction is expected to be a rare event with
711 // a heap using cms so don't do it without seeing the need.
712 for (uint i = 0; i < ParallelGCThreads; i++) {
713 _par_gc_thread_states[i]->promo.reset();
714 }
715 }
716
717 void ConcurrentMarkSweepGeneration::compute_new_size() {
718 assert_locked_or_safepoint(Heap_lock);
719
720 // If incremental collection failed, we just want to expand
721 // to the limit.
722 if (incremental_collection_failed()) {
723 clear_incremental_collection_failed();
724 grow_to_reserved();
725 return;
726 }
727
728 // The heap has been compacted but not reset yet.
729 // Any metric such as free() or used() will be incorrect.
730
731 CardGeneration::compute_new_size();
732
733 // Reset again after a possible resizing
734 if (did_compact()) {
735 cmsSpace()->reset_after_compaction();
736 }
737 }
738
739 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() {
740 assert_locked_or_safepoint(Heap_lock);
741
742 // If incremental collection failed, we just want to expand
743 // to the limit.
744 if (incremental_collection_failed()) {
745 clear_incremental_collection_failed();
746 grow_to_reserved();
747 return;
748 }
749
750 double free_percentage = ((double) free()) / capacity();
751 double desired_free_percentage = (double) MinHeapFreeRatio / 100;
752 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100;
753
754 // compute expansion delta needed for reaching desired free percentage
755 if (free_percentage < desired_free_percentage) {
756 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
757 assert(desired_capacity >= capacity(), "invalid expansion size");
758 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes);
759 Log(gc) log;
760 if (log.is_trace()) {
761 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
762 log.trace("From compute_new_size: ");
763 log.trace(" Free fraction %f", free_percentage);
764 log.trace(" Desired free fraction %f", desired_free_percentage);
765 log.trace(" Maximum free fraction %f", maximum_free_percentage);
766 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000);
767 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000);
768 CMSHeap* heap = CMSHeap::heap();
769 assert(heap->is_old_gen(this), "The CMS generation should always be the old generation");
770 size_t young_size = heap->young_gen()->capacity();
771 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000);
772 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000);
773 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000);
774 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes);
775 }
776 // safe if expansion fails
777 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio);
778 log.trace(" Expanded free fraction %f", ((double) free()) / capacity());
779 } else {
780 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage));
781 assert(desired_capacity <= capacity(), "invalid expansion size");
782 size_t shrink_bytes = capacity() - desired_capacity;
783 // Don't shrink unless the delta is greater than the minimum shrink we want
784 if (shrink_bytes >= MinHeapDeltaBytes) {
785 shrink_free_list_by(shrink_bytes);
786 }
787 }
788 }
789
790 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const {
791 return cmsSpace()->freelistLock();
792 }
793
794 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) {
795 CMSSynchronousYieldRequest yr;
796 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
797 return have_lock_and_allocate(size, tlab);
798 }
799
800 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size,
801 bool tlab /* ignored */) {
802 assert_lock_strong(freelistLock());
803 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size);
804 HeapWord* res = cmsSpace()->allocate(adjustedSize);
805 // Allocate the object live (grey) if the background collector has
806 // started marking. This is necessary because the marker may
807 // have passed this address and consequently this object will
808 // not otherwise be greyed and would be incorrectly swept up.
809 // Note that if this object contains references, the writing
810 // of those references will dirty the card containing this object
811 // allowing the object to be blackened (and its references scanned)
812 // either during a preclean phase or at the final checkpoint.
813 if (res != NULL) {
814 // We may block here with an uninitialized object with
815 // its mark-bit or P-bits not yet set. Such objects need
816 // to be safely navigable by block_start().
817 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here.");
818 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size");
819 collector()->direct_allocated(res, adjustedSize);
820 _direct_allocated_words += adjustedSize;
821 // allocation counters
822 NOT_PRODUCT(
823 _numObjectsAllocated++;
824 _numWordsAllocated += (int)adjustedSize;
825 )
826 }
827 return res;
828 }
829
830 // In the case of direct allocation by mutators in a generation that
831 // is being concurrently collected, the object must be allocated
832 // live (grey) if the background collector has started marking.
833 // This is necessary because the marker may
834 // have passed this address and consequently this object will
835 // not otherwise be greyed and would be incorrectly swept up.
836 // Note that if this object contains references, the writing
837 // of those references will dirty the card containing this object
838 // allowing the object to be blackened (and its references scanned)
839 // either during a preclean phase or at the final checkpoint.
840 void CMSCollector::direct_allocated(HeapWord* start, size_t size) {
841 assert(_markBitMap.covers(start, size), "Out of bounds");
842 if (_collectorState >= Marking) {
843 MutexLockerEx y(_markBitMap.lock(),
844 Mutex::_no_safepoint_check_flag);
845 // [see comments preceding SweepClosure::do_blk() below for details]
846 //
847 // Can the P-bits be deleted now? JJJ
848 //
849 // 1. need to mark the object as live so it isn't collected
850 // 2. need to mark the 2nd bit to indicate the object may be uninitialized
851 // 3. need to mark the end of the object so marking, precleaning or sweeping
852 // can skip over uninitialized or unparsable objects. An allocated
853 // object is considered uninitialized for our purposes as long as
854 // its klass word is NULL. All old gen objects are parsable
855 // as soon as they are initialized.)
856 _markBitMap.mark(start); // object is live
857 _markBitMap.mark(start + 1); // object is potentially uninitialized?
858 _markBitMap.mark(start + size - 1);
859 // mark end of object
860 }
861 // check that oop looks uninitialized
862 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL");
863 }
864
865 void CMSCollector::promoted(bool par, HeapWord* start,
866 bool is_obj_array, size_t obj_size) {
867 assert(_markBitMap.covers(start), "Out of bounds");
868 // See comment in direct_allocated() about when objects should
869 // be allocated live.
870 if (_collectorState >= Marking) {
871 // we already hold the marking bit map lock, taken in
872 // the prologue
873 if (par) {
874 _markBitMap.par_mark(start);
875 } else {
876 _markBitMap.mark(start);
877 }
878 // We don't need to mark the object as uninitialized (as
879 // in direct_allocated above) because this is being done with the
880 // world stopped and the object will be initialized by the
881 // time the marking, precleaning or sweeping get to look at it.
882 // But see the code for copying objects into the CMS generation,
883 // where we need to ensure that concurrent readers of the
884 // block offset table are able to safely navigate a block that
885 // is in flux from being free to being allocated (and in
886 // transition while being copied into) and subsequently
887 // becoming a bona-fide object when the copy/promotion is complete.
888 assert(SafepointSynchronize::is_at_safepoint(),
889 "expect promotion only at safepoints");
890
891 if (_collectorState < Sweeping) {
892 // Mark the appropriate cards in the modUnionTable, so that
893 // this object gets scanned before the sweep. If this is
894 // not done, CMS generation references in the object might
895 // not get marked.
896 // For the case of arrays, which are otherwise precisely
897 // marked, we need to dirty the entire array, not just its head.
898 if (is_obj_array) {
899 // The [par_]mark_range() method expects mr.end() below to
900 // be aligned to the granularity of a bit's representation
901 // in the heap. In the case of the MUT below, that's a
902 // card size.
903 MemRegion mr(start,
904 align_up(start + obj_size,
905 CardTable::card_size /* bytes */));
906 if (par) {
907 _modUnionTable.par_mark_range(mr);
908 } else {
909 _modUnionTable.mark_range(mr);
910 }
911 } else { // not an obj array; we can just mark the head
912 if (par) {
913 _modUnionTable.par_mark(start);
914 } else {
915 _modUnionTable.mark(start);
916 }
917 }
918 }
919 }
920 }
921
922 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) {
923 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in");
924 // allocate, copy and if necessary update promoinfo --
925 // delegate to underlying space.
926 assert_lock_strong(freelistLock());
927
928 #ifndef PRODUCT
929 if (CMSHeap::heap()->promotion_should_fail()) {
930 return NULL;
931 }
932 #endif // #ifndef PRODUCT
933
934 oop res = _cmsSpace->promote(obj, obj_size);
935 if (res == NULL) {
936 // expand and retry
937 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords
938 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion);
939 // Since this is the old generation, we don't try to promote
940 // into a more senior generation.
941 res = _cmsSpace->promote(obj, obj_size);
942 }
943 if (res != NULL) {
944 // See comment in allocate() about when objects should
945 // be allocated live.
946 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below");
947 collector()->promoted(false, // Not parallel
948 (HeapWord*)res, obj->is_objArray(), obj_size);
949 // promotion counters
950 NOT_PRODUCT(
951 _numObjectsPromoted++;
952 _numWordsPromoted +=
953 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size()));
954 )
955 }
956 return res;
957 }
958
959
960 // IMPORTANT: Notes on object size recognition in CMS.
961 // ---------------------------------------------------
962 // A block of storage in the CMS generation is always in
963 // one of three states. A free block (FREE), an allocated
964 // object (OBJECT) whose size() method reports the correct size,
965 // and an intermediate state (TRANSIENT) in which its size cannot
966 // be accurately determined.
967 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS)
968 // -----------------------------------------------------
969 // FREE: klass_word & 1 == 1; mark_word holds block size
970 //
971 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0;
972 // obj->size() computes correct size
973 //
974 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
975 //
976 // STATE IDENTIFICATION: (64 bit+COOPS)
977 // ------------------------------------
978 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size
979 //
980 // OBJECT: klass_word installed; klass_word != 0;
981 // obj->size() computes correct size
982 //
983 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT
984 //
985 //
986 // STATE TRANSITION DIAGRAM
987 //
988 // mut / parnew mut / parnew
989 // FREE --------------------> TRANSIENT ---------------------> OBJECT --|
990 // ^ |
991 // |------------------------ DEAD <------------------------------------|
992 // sweep mut
993 //
994 // While a block is in TRANSIENT state its size cannot be determined
995 // so readers will either need to come back later or stall until
996 // the size can be determined. Note that for the case of direct
997 // allocation, P-bits, when available, may be used to determine the
998 // size of an object that may not yet have been initialized.
999
1000 // Things to support parallel young-gen collection.
1001 oop
1002 ConcurrentMarkSweepGeneration::par_promote(int thread_num,
1003 oop old, markOop m,
1004 size_t word_sz) {
1005 #ifndef PRODUCT
1006 if (CMSHeap::heap()->promotion_should_fail()) {
1007 return NULL;
1008 }
1009 #endif // #ifndef PRODUCT
1010
1011 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1012 PromotionInfo* promoInfo = &ps->promo;
1013 // if we are tracking promotions, then first ensure space for
1014 // promotion (including spooling space for saving header if necessary).
1015 // then allocate and copy, then track promoted info if needed.
1016 // When tracking (see PromotionInfo::track()), the mark word may
1017 // be displaced and in this case restoration of the mark word
1018 // occurs in the (oop_since_save_marks_)iterate phase.
1019 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) {
1020 // Out of space for allocating spooling buffers;
1021 // try expanding and allocating spooling buffers.
1022 if (!expand_and_ensure_spooling_space(promoInfo)) {
1023 return NULL;
1024 }
1025 }
1026 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant");
1027 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz);
1028 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz);
1029 if (obj_ptr == NULL) {
1030 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz);
1031 if (obj_ptr == NULL) {
1032 return NULL;
1033 }
1034 }
1035 oop obj = oop(obj_ptr);
1036 OrderAccess::storestore();
1037 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1038 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1039 // IMPORTANT: See note on object initialization for CMS above.
1040 // Otherwise, copy the object. Here we must be careful to insert the
1041 // klass pointer last, since this marks the block as an allocated object.
1042 // Except with compressed oops it's the mark word.
1043 HeapWord* old_ptr = (HeapWord*)old;
1044 // Restore the mark word copied above.
1045 obj->set_mark(m);
1046 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1047 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1048 OrderAccess::storestore();
1049
1050 if (UseCompressedClassPointers) {
1051 // Copy gap missed by (aligned) header size calculation below
1052 obj->set_klass_gap(old->klass_gap());
1053 }
1054 if (word_sz > (size_t)oopDesc::header_size()) {
1055 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(),
1056 obj_ptr + oopDesc::header_size(),
1057 word_sz - oopDesc::header_size());
1058 }
1059
1060 // Now we can track the promoted object, if necessary. We take care
1061 // to delay the transition from uninitialized to full object
1062 // (i.e., insertion of klass pointer) until after, so that it
1063 // atomically becomes a promoted object.
1064 if (promoInfo->tracking()) {
1065 promoInfo->track((PromotedObject*)obj, old->klass());
1066 }
1067 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here.");
1068 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size");
1069 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below");
1070
1071 // Finally, install the klass pointer (this should be volatile).
1072 OrderAccess::storestore();
1073 obj->set_klass(old->klass());
1074 // We should now be able to calculate the right size for this object
1075 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object");
1076
1077 collector()->promoted(true, // parallel
1078 obj_ptr, old->is_objArray(), word_sz);
1079
1080 NOT_PRODUCT(
1081 Atomic::inc(&_numObjectsPromoted);
1082 Atomic::add(alloc_sz, &_numWordsPromoted);
1083 )
1084
1085 return obj;
1086 }
1087
1088 void
1089 ConcurrentMarkSweepGeneration::
1090 par_promote_alloc_done(int thread_num) {
1091 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1092 ps->lab.retire(thread_num);
1093 }
1094
1095 void
1096 ConcurrentMarkSweepGeneration::
1097 par_oop_since_save_marks_iterate_done(int thread_num) {
1098 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num];
1099 ParScanWithoutBarrierClosure* dummy_cl = NULL;
1100 ps->promo.promoted_oops_iterate_nv(dummy_cl);
1101
1102 // Because card-scanning has been completed, subsequent phases
1103 // (e.g., reference processing) will not need to recognize which
1104 // objects have been promoted during this GC. So, we can now disable
1105 // promotion tracking.
1106 ps->promo.stopTrackingPromotions();
1107 }
1108
1109 bool ConcurrentMarkSweepGeneration::should_collect(bool full,
1110 size_t size,
1111 bool tlab)
1112 {
1113 // We allow a STW collection only if a full
1114 // collection was requested.
1115 return full || should_allocate(size, tlab); // FIX ME !!!
1116 // This and promotion failure handling are connected at the
1117 // hip and should be fixed by untying them.
1118 }
1119
1120 bool CMSCollector::shouldConcurrentCollect() {
1121 LogTarget(Trace, gc) log;
1122
1123 if (_full_gc_requested) {
1124 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)");
1125 return true;
1126 }
1127
1128 FreelistLocker x(this);
1129 // ------------------------------------------------------------------
1130 // Print out lots of information which affects the initiation of
1131 // a collection.
1132 if (log.is_enabled() && stats().valid()) {
1133 log.print("CMSCollector shouldConcurrentCollect: ");
1134
1135 LogStream out(log);
1136 stats().print_on(&out);
1137
1138 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full());
1139 log.print("free=" SIZE_FORMAT, _cmsGen->free());
1140 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available());
1141 log.print("promotion_rate=%g", stats().promotion_rate());
1142 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate());
1143 log.print("occupancy=%3.7f", _cmsGen->occupancy());
1144 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy());
1145 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin());
1146 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end());
1147 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect());
1148 }
1149 // ------------------------------------------------------------------
1150
1151 // If the estimated time to complete a cms collection (cms_duration())
1152 // is less than the estimated time remaining until the cms generation
1153 // is full, start a collection.
1154 if (!UseCMSInitiatingOccupancyOnly) {
1155 if (stats().valid()) {
1156 if (stats().time_until_cms_start() == 0.0) {
1157 return true;
1158 }
1159 } else {
1160 // We want to conservatively collect somewhat early in order
1161 // to try and "bootstrap" our CMS/promotion statistics;
1162 // this branch will not fire after the first successful CMS
1163 // collection because the stats should then be valid.
1164 if (_cmsGen->occupancy() >= _bootstrap_occupancy) {
1165 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f",
1166 _cmsGen->occupancy(), _bootstrap_occupancy);
1167 return true;
1168 }
1169 }
1170 }
1171
1172 // Otherwise, we start a collection cycle if
1173 // old gen want a collection cycle started. Each may use
1174 // an appropriate criterion for making this decision.
1175 // XXX We need to make sure that the gen expansion
1176 // criterion dovetails well with this. XXX NEED TO FIX THIS
1177 if (_cmsGen->should_concurrent_collect()) {
1178 log.print("CMS old gen initiated");
1179 return true;
1180 }
1181
1182 // We start a collection if we believe an incremental collection may fail;
1183 // this is not likely to be productive in practice because it's probably too
1184 // late anyway.
1185 CMSHeap* heap = CMSHeap::heap();
1186 if (heap->incremental_collection_will_fail(true /* consult_young */)) {
1187 log.print("CMSCollector: collect because incremental collection will fail ");
1188 return true;
1189 }
1190
1191 if (MetaspaceGC::should_concurrent_collect()) {
1192 log.print("CMSCollector: collect for metadata allocation ");
1193 return true;
1194 }
1195
1196 // CMSTriggerInterval starts a CMS cycle if enough time has passed.
1197 if (CMSTriggerInterval >= 0) {
1198 if (CMSTriggerInterval == 0) {
1199 // Trigger always
1200 return true;
1201 }
1202
1203 // Check the CMS time since begin (we do not check the stats validity
1204 // as we want to be able to trigger the first CMS cycle as well)
1205 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) {
1206 if (stats().valid()) {
1207 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)",
1208 stats().cms_time_since_begin());
1209 } else {
1210 log.print("CMSCollector: collect because of trigger interval (first collection)");
1211 }
1212 return true;
1213 }
1214 }
1215
1216 return false;
1217 }
1218
1219 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); }
1220
1221 // Clear _expansion_cause fields of constituent generations
1222 void CMSCollector::clear_expansion_cause() {
1223 _cmsGen->clear_expansion_cause();
1224 }
1225
1226 // We should be conservative in starting a collection cycle. To
1227 // start too eagerly runs the risk of collecting too often in the
1228 // extreme. To collect too rarely falls back on full collections,
1229 // which works, even if not optimum in terms of concurrent work.
1230 // As a work around for too eagerly collecting, use the flag
1231 // UseCMSInitiatingOccupancyOnly. This also has the advantage of
1232 // giving the user an easily understandable way of controlling the
1233 // collections.
1234 // We want to start a new collection cycle if any of the following
1235 // conditions hold:
1236 // . our current occupancy exceeds the configured initiating occupancy
1237 // for this generation, or
1238 // . we recently needed to expand this space and have not, since that
1239 // expansion, done a collection of this generation, or
1240 // . the underlying space believes that it may be a good idea to initiate
1241 // a concurrent collection (this may be based on criteria such as the
1242 // following: the space uses linear allocation and linear allocation is
1243 // going to fail, or there is believed to be excessive fragmentation in
1244 // the generation, etc... or ...
1245 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for
1246 // the case of the old generation; see CR 6543076):
1247 // we may be approaching a point at which allocation requests may fail because
1248 // we will be out of sufficient free space given allocation rate estimates.]
1249 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const {
1250
1251 assert_lock_strong(freelistLock());
1252 if (occupancy() > initiating_occupancy()) {
1253 log_trace(gc)(" %s: collect because of occupancy %f / %f ",
1254 short_name(), occupancy(), initiating_occupancy());
1255 return true;
1256 }
1257 if (UseCMSInitiatingOccupancyOnly) {
1258 return false;
1259 }
1260 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) {
1261 log_trace(gc)(" %s: collect because expanded for allocation ", short_name());
1262 return true;
1263 }
1264 return false;
1265 }
1266
1267 void ConcurrentMarkSweepGeneration::collect(bool full,
1268 bool clear_all_soft_refs,
1269 size_t size,
1270 bool tlab)
1271 {
1272 collector()->collect(full, clear_all_soft_refs, size, tlab);
1273 }
1274
1275 void CMSCollector::collect(bool full,
1276 bool clear_all_soft_refs,
1277 size_t size,
1278 bool tlab)
1279 {
1280 // The following "if" branch is present for defensive reasons.
1281 // In the current uses of this interface, it can be replaced with:
1282 // assert(!GCLocker.is_active(), "Can't be called otherwise");
1283 // But I am not placing that assert here to allow future
1284 // generality in invoking this interface.
1285 if (GCLocker::is_active()) {
1286 // A consistency test for GCLocker
1287 assert(GCLocker::needs_gc(), "Should have been set already");
1288 // Skip this foreground collection, instead
1289 // expanding the heap if necessary.
1290 // Need the free list locks for the call to free() in compute_new_size()
1291 compute_new_size();
1292 return;
1293 }
1294 acquire_control_and_collect(full, clear_all_soft_refs);
1295 }
1296
1297 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) {
1298 CMSHeap* heap = CMSHeap::heap();
1299 unsigned int gc_count = heap->total_full_collections();
1300 if (gc_count == full_gc_count) {
1301 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag);
1302 _full_gc_requested = true;
1303 _full_gc_cause = cause;
1304 CGC_lock->notify(); // nudge CMS thread
1305 } else {
1306 assert(gc_count > full_gc_count, "Error: causal loop");
1307 }
1308 }
1309
1310 bool CMSCollector::is_external_interruption() {
1311 GCCause::Cause cause = CMSHeap::heap()->gc_cause();
1312 return GCCause::is_user_requested_gc(cause) ||
1313 GCCause::is_serviceability_requested_gc(cause);
1314 }
1315
1316 void CMSCollector::report_concurrent_mode_interruption() {
1317 if (is_external_interruption()) {
1318 log_debug(gc)("Concurrent mode interrupted");
1319 } else {
1320 log_debug(gc)("Concurrent mode failure");
1321 _gc_tracer_cm->report_concurrent_mode_failure();
1322 }
1323 }
1324
1325
1326 // The foreground and background collectors need to coordinate in order
1327 // to make sure that they do not mutually interfere with CMS collections.
1328 // When a background collection is active,
1329 // the foreground collector may need to take over (preempt) and
1330 // synchronously complete an ongoing collection. Depending on the
1331 // frequency of the background collections and the heap usage
1332 // of the application, this preemption can be seldom or frequent.
1333 // There are only certain
1334 // points in the background collection that the "collection-baton"
1335 // can be passed to the foreground collector.
1336 //
1337 // The foreground collector will wait for the baton before
1338 // starting any part of the collection. The foreground collector
1339 // will only wait at one location.
1340 //
1341 // The background collector will yield the baton before starting a new
1342 // phase of the collection (e.g., before initial marking, marking from roots,
1343 // precleaning, final re-mark, sweep etc.) This is normally done at the head
1344 // of the loop which switches the phases. The background collector does some
1345 // of the phases (initial mark, final re-mark) with the world stopped.
1346 // Because of locking involved in stopping the world,
1347 // the foreground collector should not block waiting for the background
1348 // collector when it is doing a stop-the-world phase. The background
1349 // collector will yield the baton at an additional point just before
1350 // it enters a stop-the-world phase. Once the world is stopped, the
1351 // background collector checks the phase of the collection. If the
1352 // phase has not changed, it proceeds with the collection. If the
1353 // phase has changed, it skips that phase of the collection. See
1354 // the comments on the use of the Heap_lock in collect_in_background().
1355 //
1356 // Variable used in baton passing.
1357 // _foregroundGCIsActive - Set to true by the foreground collector when
1358 // it wants the baton. The foreground clears it when it has finished
1359 // the collection.
1360 // _foregroundGCShouldWait - Set to true by the background collector
1361 // when it is running. The foreground collector waits while
1362 // _foregroundGCShouldWait is true.
1363 // CGC_lock - monitor used to protect access to the above variables
1364 // and to notify the foreground and background collectors.
1365 // _collectorState - current state of the CMS collection.
1366 //
1367 // The foreground collector
1368 // acquires the CGC_lock
1369 // sets _foregroundGCIsActive
1370 // waits on the CGC_lock for _foregroundGCShouldWait to be false
1371 // various locks acquired in preparation for the collection
1372 // are released so as not to block the background collector
1373 // that is in the midst of a collection
1374 // proceeds with the collection
1375 // clears _foregroundGCIsActive
1376 // returns
1377 //
1378 // The background collector in a loop iterating on the phases of the
1379 // collection
1380 // acquires the CGC_lock
1381 // sets _foregroundGCShouldWait
1382 // if _foregroundGCIsActive is set
1383 // clears _foregroundGCShouldWait, notifies _CGC_lock
1384 // waits on _CGC_lock for _foregroundGCIsActive to become false
1385 // and exits the loop.
1386 // otherwise
1387 // proceed with that phase of the collection
1388 // if the phase is a stop-the-world phase,
1389 // yield the baton once more just before enqueueing
1390 // the stop-world CMS operation (executed by the VM thread).
1391 // returns after all phases of the collection are done
1392 //
1393
1394 void CMSCollector::acquire_control_and_collect(bool full,
1395 bool clear_all_soft_refs) {
1396 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
1397 assert(!Thread::current()->is_ConcurrentGC_thread(),
1398 "shouldn't try to acquire control from self!");
1399
1400 // Start the protocol for acquiring control of the
1401 // collection from the background collector (aka CMS thread).
1402 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1403 "VM thread should have CMS token");
1404 // Remember the possibly interrupted state of an ongoing
1405 // concurrent collection
1406 CollectorState first_state = _collectorState;
1407
1408 // Signal to a possibly ongoing concurrent collection that
1409 // we want to do a foreground collection.
1410 _foregroundGCIsActive = true;
1411
1412 // release locks and wait for a notify from the background collector
1413 // releasing the locks in only necessary for phases which
1414 // do yields to improve the granularity of the collection.
1415 assert_lock_strong(bitMapLock());
1416 // We need to lock the Free list lock for the space that we are
1417 // currently collecting.
1418 assert(haveFreelistLocks(), "Must be holding free list locks");
1419 bitMapLock()->unlock();
1420 releaseFreelistLocks();
1421 {
1422 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1423 if (_foregroundGCShouldWait) {
1424 // We are going to be waiting for action for the CMS thread;
1425 // it had better not be gone (for instance at shutdown)!
1426 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(),
1427 "CMS thread must be running");
1428 // Wait here until the background collector gives us the go-ahead
1429 ConcurrentMarkSweepThread::clear_CMS_flag(
1430 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token
1431 // Get a possibly blocked CMS thread going:
1432 // Note that we set _foregroundGCIsActive true above,
1433 // without protection of the CGC_lock.
1434 CGC_lock->notify();
1435 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(),
1436 "Possible deadlock");
1437 while (_foregroundGCShouldWait) {
1438 // wait for notification
1439 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1440 // Possibility of delay/starvation here, since CMS token does
1441 // not know to give priority to VM thread? Actually, i think
1442 // there wouldn't be any delay/starvation, but the proof of
1443 // that "fact" (?) appears non-trivial. XXX 20011219YSR
1444 }
1445 ConcurrentMarkSweepThread::set_CMS_flag(
1446 ConcurrentMarkSweepThread::CMS_vm_has_token);
1447 }
1448 }
1449 // The CMS_token is already held. Get back the other locks.
1450 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(),
1451 "VM thread should have CMS token");
1452 getFreelistLocks();
1453 bitMapLock()->lock_without_safepoint_check();
1454 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d",
1455 p2i(Thread::current()), first_state);
1456 log_debug(gc, state)(" gets control with state %d", _collectorState);
1457
1458 // Inform cms gen if this was due to partial collection failing.
1459 // The CMS gen may use this fact to determine its expansion policy.
1460 CMSHeap* heap = CMSHeap::heap();
1461 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) {
1462 assert(!_cmsGen->incremental_collection_failed(),
1463 "Should have been noticed, reacted to and cleared");
1464 _cmsGen->set_incremental_collection_failed();
1465 }
1466
1467 if (first_state > Idling) {
1468 report_concurrent_mode_interruption();
1469 }
1470
1471 set_did_compact(true);
1472
1473 // If the collection is being acquired from the background
1474 // collector, there may be references on the discovered
1475 // references lists. Abandon those references, since some
1476 // of them may have become unreachable after concurrent
1477 // discovery; the STW compacting collector will redo discovery
1478 // more precisely, without being subject to floating garbage.
1479 // Leaving otherwise unreachable references in the discovered
1480 // lists would require special handling.
1481 ref_processor()->disable_discovery();
1482 ref_processor()->abandon_partial_discovery();
1483 ref_processor()->verify_no_references_recorded();
1484
1485 if (first_state > Idling) {
1486 save_heap_summary();
1487 }
1488
1489 do_compaction_work(clear_all_soft_refs);
1490
1491 // Has the GC time limit been exceeded?
1492 size_t max_eden_size = _young_gen->max_eden_size();
1493 GCCause::Cause gc_cause = heap->gc_cause();
1494 size_policy()->check_gc_overhead_limit(_young_gen->used(),
1495 _young_gen->eden()->used(),
1496 _cmsGen->max_capacity(),
1497 max_eden_size,
1498 full,
1499 gc_cause,
1500 heap->soft_ref_policy());
1501
1502 // Reset the expansion cause, now that we just completed
1503 // a collection cycle.
1504 clear_expansion_cause();
1505 _foregroundGCIsActive = false;
1506 return;
1507 }
1508
1509 // Resize the tenured generation
1510 // after obtaining the free list locks for the
1511 // two generations.
1512 void CMSCollector::compute_new_size() {
1513 assert_locked_or_safepoint(Heap_lock);
1514 FreelistLocker z(this);
1515 MetaspaceGC::compute_new_size();
1516 _cmsGen->compute_new_size_free_list();
1517 }
1518
1519 // A work method used by the foreground collector to do
1520 // a mark-sweep-compact.
1521 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) {
1522 CMSHeap* heap = CMSHeap::heap();
1523
1524 STWGCTimer* gc_timer = GenMarkSweep::gc_timer();
1525 gc_timer->register_gc_start();
1526
1527 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer();
1528 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start());
1529
1530 heap->pre_full_gc_dump(gc_timer);
1531
1532 GCTraceTime(Trace, gc, phases) t("CMS:MSC");
1533
1534 // Temporarily widen the span of the weak reference processing to
1535 // the entire heap.
1536 MemRegion new_span(CMSHeap::heap()->reserved_region());
1537 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span);
1538 // Temporarily, clear the "is_alive_non_header" field of the
1539 // reference processor.
1540 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL);
1541 // Temporarily make reference _processing_ single threaded (non-MT).
1542 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false);
1543 // Temporarily make refs discovery atomic
1544 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true);
1545 // Temporarily make reference _discovery_ single threaded (non-MT)
1546 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
1547
1548 ref_processor()->set_enqueuing_is_done(false);
1549 ref_processor()->enable_discovery();
1550 ref_processor()->setup_policy(clear_all_soft_refs);
1551 // If an asynchronous collection finishes, the _modUnionTable is
1552 // all clear. If we are assuming the collection from an asynchronous
1553 // collection, clear the _modUnionTable.
1554 assert(_collectorState != Idling || _modUnionTable.isAllClear(),
1555 "_modUnionTable should be clear if the baton was not passed");
1556 _modUnionTable.clear_all();
1557 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(),
1558 "mod union for klasses should be clear if the baton was passed");
1559 _ct->cld_rem_set()->clear_mod_union();
1560
1561
1562 // We must adjust the allocation statistics being maintained
1563 // in the free list space. We do so by reading and clearing
1564 // the sweep timer and updating the block flux rate estimates below.
1565 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive");
1566 if (_inter_sweep_timer.is_active()) {
1567 _inter_sweep_timer.stop();
1568 // Note that we do not use this sample to update the _inter_sweep_estimate.
1569 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
1570 _inter_sweep_estimate.padded_average(),
1571 _intra_sweep_estimate.padded_average());
1572 }
1573
1574 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs);
1575 #ifdef ASSERT
1576 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
1577 size_t free_size = cms_space->free();
1578 assert(free_size ==
1579 pointer_delta(cms_space->end(), cms_space->compaction_top())
1580 * HeapWordSize,
1581 "All the free space should be compacted into one chunk at top");
1582 assert(cms_space->dictionary()->total_chunk_size(
1583 debug_only(cms_space->freelistLock())) == 0 ||
1584 cms_space->totalSizeInIndexedFreeLists() == 0,
1585 "All the free space should be in a single chunk");
1586 size_t num = cms_space->totalCount();
1587 assert((free_size == 0 && num == 0) ||
1588 (free_size > 0 && (num == 1 || num == 2)),
1589 "There should be at most 2 free chunks after compaction");
1590 #endif // ASSERT
1591 _collectorState = Resetting;
1592 assert(_restart_addr == NULL,
1593 "Should have been NULL'd before baton was passed");
1594 reset_stw();
1595 _cmsGen->reset_after_compaction();
1596 _concurrent_cycles_since_last_unload = 0;
1597
1598 // Clear any data recorded in the PLAB chunk arrays.
1599 if (_survivor_plab_array != NULL) {
1600 reset_survivor_plab_arrays();
1601 }
1602
1603 // Adjust the per-size allocation stats for the next epoch.
1604 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */);
1605 // Restart the "inter sweep timer" for the next epoch.
1606 _inter_sweep_timer.reset();
1607 _inter_sweep_timer.start();
1608
1609 // No longer a need to do a concurrent collection for Metaspace.
1610 MetaspaceGC::set_should_concurrent_collect(false);
1611
1612 heap->post_full_gc_dump(gc_timer);
1613
1614 gc_timer->register_gc_end();
1615
1616 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions());
1617
1618 // For a mark-sweep-compact, compute_new_size() will be called
1619 // in the heap's do_collection() method.
1620 }
1621
1622 void CMSCollector::print_eden_and_survivor_chunk_arrays() {
1623 Log(gc, heap) log;
1624 if (!log.is_trace()) {
1625 return;
1626 }
1627
1628 ContiguousSpace* eden_space = _young_gen->eden();
1629 ContiguousSpace* from_space = _young_gen->from();
1630 ContiguousSpace* to_space = _young_gen->to();
1631 // Eden
1632 if (_eden_chunk_array != NULL) {
1633 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1634 p2i(eden_space->bottom()), p2i(eden_space->top()),
1635 p2i(eden_space->end()), eden_space->capacity());
1636 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT,
1637 _eden_chunk_index, _eden_chunk_capacity);
1638 for (size_t i = 0; i < _eden_chunk_index; i++) {
1639 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i]));
1640 }
1641 }
1642 // Survivor
1643 if (_survivor_chunk_array != NULL) {
1644 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")",
1645 p2i(from_space->bottom()), p2i(from_space->top()),
1646 p2i(from_space->end()), from_space->capacity());
1647 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT,
1648 _survivor_chunk_index, _survivor_chunk_capacity);
1649 for (size_t i = 0; i < _survivor_chunk_index; i++) {
1650 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i]));
1651 }
1652 }
1653 }
1654
1655 void CMSCollector::getFreelistLocks() const {
1656 // Get locks for all free lists in all generations that this
1657 // collector is responsible for
1658 _cmsGen->freelistLock()->lock_without_safepoint_check();
1659 }
1660
1661 void CMSCollector::releaseFreelistLocks() const {
1662 // Release locks for all free lists in all generations that this
1663 // collector is responsible for
1664 _cmsGen->freelistLock()->unlock();
1665 }
1666
1667 bool CMSCollector::haveFreelistLocks() const {
1668 // Check locks for all free lists in all generations that this
1669 // collector is responsible for
1670 assert_lock_strong(_cmsGen->freelistLock());
1671 PRODUCT_ONLY(ShouldNotReachHere());
1672 return true;
1673 }
1674
1675 // A utility class that is used by the CMS collector to
1676 // temporarily "release" the foreground collector from its
1677 // usual obligation to wait for the background collector to
1678 // complete an ongoing phase before proceeding.
1679 class ReleaseForegroundGC: public StackObj {
1680 private:
1681 CMSCollector* _c;
1682 public:
1683 ReleaseForegroundGC(CMSCollector* c) : _c(c) {
1684 assert(_c->_foregroundGCShouldWait, "Else should not need to call");
1685 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1686 // allow a potentially blocked foreground collector to proceed
1687 _c->_foregroundGCShouldWait = false;
1688 if (_c->_foregroundGCIsActive) {
1689 CGC_lock->notify();
1690 }
1691 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1692 "Possible deadlock");
1693 }
1694
1695 ~ReleaseForegroundGC() {
1696 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?");
1697 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1698 _c->_foregroundGCShouldWait = true;
1699 }
1700 };
1701
1702 void CMSCollector::collect_in_background(GCCause::Cause cause) {
1703 assert(Thread::current()->is_ConcurrentGC_thread(),
1704 "A CMS asynchronous collection is only allowed on a CMS thread.");
1705
1706 CMSHeap* heap = CMSHeap::heap();
1707 {
1708 bool safepoint_check = Mutex::_no_safepoint_check_flag;
1709 MutexLockerEx hl(Heap_lock, safepoint_check);
1710 FreelistLocker fll(this);
1711 MutexLockerEx x(CGC_lock, safepoint_check);
1712 if (_foregroundGCIsActive) {
1713 // The foreground collector is. Skip this
1714 // background collection.
1715 assert(!_foregroundGCShouldWait, "Should be clear");
1716 return;
1717 } else {
1718 assert(_collectorState == Idling, "Should be idling before start.");
1719 _collectorState = InitialMarking;
1720 register_gc_start(cause);
1721 // Reset the expansion cause, now that we are about to begin
1722 // a new cycle.
1723 clear_expansion_cause();
1724
1725 // Clear the MetaspaceGC flag since a concurrent collection
1726 // is starting but also clear it after the collection.
1727 MetaspaceGC::set_should_concurrent_collect(false);
1728 }
1729 // Decide if we want to enable class unloading as part of the
1730 // ensuing concurrent GC cycle.
1731 update_should_unload_classes();
1732 _full_gc_requested = false; // acks all outstanding full gc requests
1733 _full_gc_cause = GCCause::_no_gc;
1734 // Signal that we are about to start a collection
1735 heap->increment_total_full_collections(); // ... starting a collection cycle
1736 _collection_count_start = heap->total_full_collections();
1737 }
1738
1739 size_t prev_used = _cmsGen->used();
1740
1741 // The change of the collection state is normally done at this level;
1742 // the exceptions are phases that are executed while the world is
1743 // stopped. For those phases the change of state is done while the
1744 // world is stopped. For baton passing purposes this allows the
1745 // background collector to finish the phase and change state atomically.
1746 // The foreground collector cannot wait on a phase that is done
1747 // while the world is stopped because the foreground collector already
1748 // has the world stopped and would deadlock.
1749 while (_collectorState != Idling) {
1750 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d",
1751 p2i(Thread::current()), _collectorState);
1752 // The foreground collector
1753 // holds the Heap_lock throughout its collection.
1754 // holds the CMS token (but not the lock)
1755 // except while it is waiting for the background collector to yield.
1756 //
1757 // The foreground collector should be blocked (not for long)
1758 // if the background collector is about to start a phase
1759 // executed with world stopped. If the background
1760 // collector has already started such a phase, the
1761 // foreground collector is blocked waiting for the
1762 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking)
1763 // are executed in the VM thread.
1764 //
1765 // The locking order is
1766 // PendingListLock (PLL) -- if applicable (FinalMarking)
1767 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue())
1768 // CMS token (claimed in
1769 // stop_world_and_do() -->
1770 // safepoint_synchronize() -->
1771 // CMSThread::synchronize())
1772
1773 {
1774 // Check if the FG collector wants us to yield.
1775 CMSTokenSync x(true); // is cms thread
1776 if (waitForForegroundGC()) {
1777 // We yielded to a foreground GC, nothing more to be
1778 // done this round.
1779 assert(_foregroundGCShouldWait == false, "We set it to false in "
1780 "waitForForegroundGC()");
1781 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1782 p2i(Thread::current()), _collectorState);
1783 return;
1784 } else {
1785 // The background collector can run but check to see if the
1786 // foreground collector has done a collection while the
1787 // background collector was waiting to get the CGC_lock
1788 // above. If yes, break so that _foregroundGCShouldWait
1789 // is cleared before returning.
1790 if (_collectorState == Idling) {
1791 break;
1792 }
1793 }
1794 }
1795
1796 assert(_foregroundGCShouldWait, "Foreground collector, if active, "
1797 "should be waiting");
1798
1799 switch (_collectorState) {
1800 case InitialMarking:
1801 {
1802 ReleaseForegroundGC x(this);
1803 stats().record_cms_begin();
1804 VM_CMS_Initial_Mark initial_mark_op(this);
1805 VMThread::execute(&initial_mark_op);
1806 }
1807 // The collector state may be any legal state at this point
1808 // since the background collector may have yielded to the
1809 // foreground collector.
1810 break;
1811 case Marking:
1812 // initial marking in checkpointRootsInitialWork has been completed
1813 if (markFromRoots()) { // we were successful
1814 assert(_collectorState == Precleaning, "Collector state should "
1815 "have changed");
1816 } else {
1817 assert(_foregroundGCIsActive, "Internal state inconsistency");
1818 }
1819 break;
1820 case Precleaning:
1821 // marking from roots in markFromRoots has been completed
1822 preclean();
1823 assert(_collectorState == AbortablePreclean ||
1824 _collectorState == FinalMarking,
1825 "Collector state should have changed");
1826 break;
1827 case AbortablePreclean:
1828 abortable_preclean();
1829 assert(_collectorState == FinalMarking, "Collector state should "
1830 "have changed");
1831 break;
1832 case FinalMarking:
1833 {
1834 ReleaseForegroundGC x(this);
1835
1836 VM_CMS_Final_Remark final_remark_op(this);
1837 VMThread::execute(&final_remark_op);
1838 }
1839 assert(_foregroundGCShouldWait, "block post-condition");
1840 break;
1841 case Sweeping:
1842 // final marking in checkpointRootsFinal has been completed
1843 sweep();
1844 assert(_collectorState == Resizing, "Collector state change "
1845 "to Resizing must be done under the free_list_lock");
1846
1847 case Resizing: {
1848 // Sweeping has been completed...
1849 // At this point the background collection has completed.
1850 // Don't move the call to compute_new_size() down
1851 // into code that might be executed if the background
1852 // collection was preempted.
1853 {
1854 ReleaseForegroundGC x(this); // unblock FG collection
1855 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag);
1856 CMSTokenSync z(true); // not strictly needed.
1857 if (_collectorState == Resizing) {
1858 compute_new_size();
1859 save_heap_summary();
1860 _collectorState = Resetting;
1861 } else {
1862 assert(_collectorState == Idling, "The state should only change"
1863 " because the foreground collector has finished the collection");
1864 }
1865 }
1866 break;
1867 }
1868 case Resetting:
1869 // CMS heap resizing has been completed
1870 reset_concurrent();
1871 assert(_collectorState == Idling, "Collector state should "
1872 "have changed");
1873
1874 MetaspaceGC::set_should_concurrent_collect(false);
1875
1876 stats().record_cms_end();
1877 // Don't move the concurrent_phases_end() and compute_new_size()
1878 // calls to here because a preempted background collection
1879 // has it's state set to "Resetting".
1880 break;
1881 case Idling:
1882 default:
1883 ShouldNotReachHere();
1884 break;
1885 }
1886 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d",
1887 p2i(Thread::current()), _collectorState);
1888 assert(_foregroundGCShouldWait, "block post-condition");
1889 }
1890
1891 // Should this be in gc_epilogue?
1892 heap->counters()->update_counters();
1893
1894 {
1895 // Clear _foregroundGCShouldWait and, in the event that the
1896 // foreground collector is waiting, notify it, before
1897 // returning.
1898 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1899 _foregroundGCShouldWait = false;
1900 if (_foregroundGCIsActive) {
1901 CGC_lock->notify();
1902 }
1903 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1904 "Possible deadlock");
1905 }
1906 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d",
1907 p2i(Thread::current()), _collectorState);
1908 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)",
1909 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K);
1910 }
1911
1912 void CMSCollector::register_gc_start(GCCause::Cause cause) {
1913 _cms_start_registered = true;
1914 _gc_timer_cm->register_gc_start();
1915 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start());
1916 }
1917
1918 void CMSCollector::register_gc_end() {
1919 if (_cms_start_registered) {
1920 report_heap_summary(GCWhen::AfterGC);
1921
1922 _gc_timer_cm->register_gc_end();
1923 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1924 _cms_start_registered = false;
1925 }
1926 }
1927
1928 void CMSCollector::save_heap_summary() {
1929 CMSHeap* heap = CMSHeap::heap();
1930 _last_heap_summary = heap->create_heap_summary();
1931 _last_metaspace_summary = heap->create_metaspace_summary();
1932 }
1933
1934 void CMSCollector::report_heap_summary(GCWhen::Type when) {
1935 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary);
1936 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary);
1937 }
1938
1939 bool CMSCollector::waitForForegroundGC() {
1940 bool res = false;
1941 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
1942 "CMS thread should have CMS token");
1943 // Block the foreground collector until the
1944 // background collectors decides whether to
1945 // yield.
1946 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
1947 _foregroundGCShouldWait = true;
1948 if (_foregroundGCIsActive) {
1949 // The background collector yields to the
1950 // foreground collector and returns a value
1951 // indicating that it has yielded. The foreground
1952 // collector can proceed.
1953 res = true;
1954 _foregroundGCShouldWait = false;
1955 ConcurrentMarkSweepThread::clear_CMS_flag(
1956 ConcurrentMarkSweepThread::CMS_cms_has_token);
1957 ConcurrentMarkSweepThread::set_CMS_flag(
1958 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1959 // Get a possibly blocked foreground thread going
1960 CGC_lock->notify();
1961 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d",
1962 p2i(Thread::current()), _collectorState);
1963 while (_foregroundGCIsActive) {
1964 CGC_lock->wait(Mutex::_no_safepoint_check_flag);
1965 }
1966 ConcurrentMarkSweepThread::set_CMS_flag(
1967 ConcurrentMarkSweepThread::CMS_cms_has_token);
1968 ConcurrentMarkSweepThread::clear_CMS_flag(
1969 ConcurrentMarkSweepThread::CMS_cms_wants_token);
1970 }
1971 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d",
1972 p2i(Thread::current()), _collectorState);
1973 return res;
1974 }
1975
1976 // Because of the need to lock the free lists and other structures in
1977 // the collector, common to all the generations that the collector is
1978 // collecting, we need the gc_prologues of individual CMS generations
1979 // delegate to their collector. It may have been simpler had the
1980 // current infrastructure allowed one to call a prologue on a
1981 // collector. In the absence of that we have the generation's
1982 // prologue delegate to the collector, which delegates back
1983 // some "local" work to a worker method in the individual generations
1984 // that it's responsible for collecting, while itself doing any
1985 // work common to all generations it's responsible for. A similar
1986 // comment applies to the gc_epilogue()'s.
1987 // The role of the variable _between_prologue_and_epilogue is to
1988 // enforce the invocation protocol.
1989 void CMSCollector::gc_prologue(bool full) {
1990 // Call gc_prologue_work() for the CMSGen
1991 // we are responsible for.
1992
1993 // The following locking discipline assumes that we are only called
1994 // when the world is stopped.
1995 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption");
1996
1997 // The CMSCollector prologue must call the gc_prologues for the
1998 // "generations" that it's responsible
1999 // for.
2000
2001 assert( Thread::current()->is_VM_thread()
2002 || ( CMSScavengeBeforeRemark
2003 && Thread::current()->is_ConcurrentGC_thread()),
2004 "Incorrect thread type for prologue execution");
2005
2006 if (_between_prologue_and_epilogue) {
2007 // We have already been invoked; this is a gc_prologue delegation
2008 // from yet another CMS generation that we are responsible for, just
2009 // ignore it since all relevant work has already been done.
2010 return;
2011 }
2012
2013 // set a bit saying prologue has been called; cleared in epilogue
2014 _between_prologue_and_epilogue = true;
2015 // Claim locks for common data structures, then call gc_prologue_work()
2016 // for each CMSGen.
2017
2018 getFreelistLocks(); // gets free list locks on constituent spaces
2019 bitMapLock()->lock_without_safepoint_check();
2020
2021 // Should call gc_prologue_work() for all cms gens we are responsible for
2022 bool duringMarking = _collectorState >= Marking
2023 && _collectorState < Sweeping;
2024
2025 // The young collections clear the modified oops state, which tells if
2026 // there are any modified oops in the class. The remark phase also needs
2027 // that information. Tell the young collection to save the union of all
2028 // modified klasses.
2029 if (duringMarking) {
2030 _ct->cld_rem_set()->set_accumulate_modified_oops(true);
2031 }
2032
2033 bool registerClosure = duringMarking;
2034
2035 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar);
2036
2037 if (!full) {
2038 stats().record_gc0_begin();
2039 }
2040 }
2041
2042 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) {
2043
2044 _capacity_at_prologue = capacity();
2045 _used_at_prologue = used();
2046
2047 // We enable promotion tracking so that card-scanning can recognize
2048 // which objects have been promoted during this GC and skip them.
2049 for (uint i = 0; i < ParallelGCThreads; i++) {
2050 _par_gc_thread_states[i]->promo.startTrackingPromotions();
2051 }
2052
2053 // Delegate to CMScollector which knows how to coordinate between
2054 // this and any other CMS generations that it is responsible for
2055 // collecting.
2056 collector()->gc_prologue(full);
2057 }
2058
2059 // This is a "private" interface for use by this generation's CMSCollector.
2060 // Not to be called directly by any other entity (for instance,
2061 // GenCollectedHeap, which calls the "public" gc_prologue method above).
2062 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full,
2063 bool registerClosure, ModUnionClosure* modUnionClosure) {
2064 assert(!incremental_collection_failed(), "Shouldn't be set yet");
2065 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL,
2066 "Should be NULL");
2067 if (registerClosure) {
2068 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure);
2069 }
2070 cmsSpace()->gc_prologue();
2071 // Clear stat counters
2072 NOT_PRODUCT(
2073 assert(_numObjectsPromoted == 0, "check");
2074 assert(_numWordsPromoted == 0, "check");
2075 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently",
2076 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord));
2077 _numObjectsAllocated = 0;
2078 _numWordsAllocated = 0;
2079 )
2080 }
2081
2082 void CMSCollector::gc_epilogue(bool full) {
2083 // The following locking discipline assumes that we are only called
2084 // when the world is stopped.
2085 assert(SafepointSynchronize::is_at_safepoint(),
2086 "world is stopped assumption");
2087
2088 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks
2089 // if linear allocation blocks need to be appropriately marked to allow the
2090 // the blocks to be parsable. We also check here whether we need to nudge the
2091 // CMS collector thread to start a new cycle (if it's not already active).
2092 assert( Thread::current()->is_VM_thread()
2093 || ( CMSScavengeBeforeRemark
2094 && Thread::current()->is_ConcurrentGC_thread()),
2095 "Incorrect thread type for epilogue execution");
2096
2097 if (!_between_prologue_and_epilogue) {
2098 // We have already been invoked; this is a gc_epilogue delegation
2099 // from yet another CMS generation that we are responsible for, just
2100 // ignore it since all relevant work has already been done.
2101 return;
2102 }
2103 assert(haveFreelistLocks(), "must have freelist locks");
2104 assert_lock_strong(bitMapLock());
2105
2106 _ct->cld_rem_set()->set_accumulate_modified_oops(false);
2107
2108 _cmsGen->gc_epilogue_work(full);
2109
2110 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) {
2111 // in case sampling was not already enabled, enable it
2112 _start_sampling = true;
2113 }
2114 // reset _eden_chunk_array so sampling starts afresh
2115 _eden_chunk_index = 0;
2116
2117 size_t cms_used = _cmsGen->cmsSpace()->used();
2118
2119 // update performance counters - this uses a special version of
2120 // update_counters() that allows the utilization to be passed as a
2121 // parameter, avoiding multiple calls to used().
2122 //
2123 _cmsGen->update_counters(cms_used);
2124
2125 bitMapLock()->unlock();
2126 releaseFreelistLocks();
2127
2128 if (!CleanChunkPoolAsync) {
2129 Chunk::clean_chunk_pool();
2130 }
2131
2132 set_did_compact(false);
2133 _between_prologue_and_epilogue = false; // ready for next cycle
2134 }
2135
2136 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) {
2137 collector()->gc_epilogue(full);
2138
2139 // When using ParNew, promotion tracking should have already been
2140 // disabled. However, the prologue (which enables promotion
2141 // tracking) and epilogue are called irrespective of the type of
2142 // GC. So they will also be called before and after Full GCs, during
2143 // which promotion tracking will not be explicitly disabled. So,
2144 // it's safer to also disable it here too (to be symmetric with
2145 // enabling it in the prologue).
2146 for (uint i = 0; i < ParallelGCThreads; i++) {
2147 _par_gc_thread_states[i]->promo.stopTrackingPromotions();
2148 }
2149 }
2150
2151 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) {
2152 assert(!incremental_collection_failed(), "Should have been cleared");
2153 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL);
2154 cmsSpace()->gc_epilogue();
2155 // Print stat counters
2156 NOT_PRODUCT(
2157 assert(_numObjectsAllocated == 0, "check");
2158 assert(_numWordsAllocated == 0, "check");
2159 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
2160 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord));
2161 _numObjectsPromoted = 0;
2162 _numWordsPromoted = 0;
2163 )
2164
2165 // Call down the chain in contiguous_available needs the freelistLock
2166 // so print this out before releasing the freeListLock.
2167 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available());
2168 }
2169
2170 #ifndef PRODUCT
2171 bool CMSCollector::have_cms_token() {
2172 Thread* thr = Thread::current();
2173 if (thr->is_VM_thread()) {
2174 return ConcurrentMarkSweepThread::vm_thread_has_cms_token();
2175 } else if (thr->is_ConcurrentGC_thread()) {
2176 return ConcurrentMarkSweepThread::cms_thread_has_cms_token();
2177 } else if (thr->is_GC_task_thread()) {
2178 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() &&
2179 ParGCRareEvent_lock->owned_by_self();
2180 }
2181 return false;
2182 }
2183
2184 // Check reachability of the given heap address in CMS generation,
2185 // treating all other generations as roots.
2186 bool CMSCollector::is_cms_reachable(HeapWord* addr) {
2187 // We could "guarantee" below, rather than assert, but I'll
2188 // leave these as "asserts" so that an adventurous debugger
2189 // could try this in the product build provided some subset of
2190 // the conditions were met, provided they were interested in the
2191 // results and knew that the computation below wouldn't interfere
2192 // with other concurrent computations mutating the structures
2193 // being read or written.
2194 assert(SafepointSynchronize::is_at_safepoint(),
2195 "Else mutations in object graph will make answer suspect");
2196 assert(have_cms_token(), "Should hold cms token");
2197 assert(haveFreelistLocks(), "must hold free list locks");
2198 assert_lock_strong(bitMapLock());
2199
2200 // Clear the marking bit map array before starting, but, just
2201 // for kicks, first report if the given address is already marked
2202 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr),
2203 _markBitMap.isMarked(addr) ? "" : " not");
2204
2205 if (verify_after_remark()) {
2206 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2207 bool result = verification_mark_bm()->isMarked(addr);
2208 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr),
2209 result ? "IS" : "is NOT");
2210 return result;
2211 } else {
2212 tty->print_cr("Could not compute result");
2213 return false;
2214 }
2215 }
2216 #endif
2217
2218 void
2219 CMSCollector::print_on_error(outputStream* st) {
2220 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector;
2221 if (collector != NULL) {
2222 CMSBitMap* bitmap = &collector->_markBitMap;
2223 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap));
2224 bitmap->print_on_error(st, " Bits: ");
2225
2226 st->cr();
2227
2228 CMSBitMap* mut_bitmap = &collector->_modUnionTable;
2229 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap));
2230 mut_bitmap->print_on_error(st, " Bits: ");
2231 }
2232 }
2233
2234 ////////////////////////////////////////////////////////
2235 // CMS Verification Support
2236 ////////////////////////////////////////////////////////
2237 // Following the remark phase, the following invariant
2238 // should hold -- each object in the CMS heap which is
2239 // marked in markBitMap() should be marked in the verification_mark_bm().
2240
2241 class VerifyMarkedClosure: public BitMapClosure {
2242 CMSBitMap* _marks;
2243 bool _failed;
2244
2245 public:
2246 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {}
2247
2248 bool do_bit(size_t offset) {
2249 HeapWord* addr = _marks->offsetToHeapWord(offset);
2250 if (!_marks->isMarked(addr)) {
2251 Log(gc, verify) log;
2252 ResourceMark rm;
2253 LogStream ls(log.error());
2254 oop(addr)->print_on(&ls);
2255 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
2256 _failed = true;
2257 }
2258 return true;
2259 }
2260
2261 bool failed() { return _failed; }
2262 };
2263
2264 bool CMSCollector::verify_after_remark() {
2265 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking.");
2266 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag);
2267 static bool init = false;
2268
2269 assert(SafepointSynchronize::is_at_safepoint(),
2270 "Else mutations in object graph will make answer suspect");
2271 assert(have_cms_token(),
2272 "Else there may be mutual interference in use of "
2273 " verification data structures");
2274 assert(_collectorState > Marking && _collectorState <= Sweeping,
2275 "Else marking info checked here may be obsolete");
2276 assert(haveFreelistLocks(), "must hold free list locks");
2277 assert_lock_strong(bitMapLock());
2278
2279
2280 // Allocate marking bit map if not already allocated
2281 if (!init) { // first time
2282 if (!verification_mark_bm()->allocate(_span)) {
2283 return false;
2284 }
2285 init = true;
2286 }
2287
2288 assert(verification_mark_stack()->isEmpty(), "Should be empty");
2289
2290 // Turn off refs discovery -- so we will be tracing through refs.
2291 // This is as intended, because by this time
2292 // GC must already have cleared any refs that need to be cleared,
2293 // and traced those that need to be marked; moreover,
2294 // the marking done here is not going to interfere in any
2295 // way with the marking information used by GC.
2296 NoRefDiscovery no_discovery(ref_processor());
2297
2298 #if COMPILER2_OR_JVMCI
2299 DerivedPointerTableDeactivate dpt_deact;
2300 #endif
2301
2302 // Clear any marks from a previous round
2303 verification_mark_bm()->clear_all();
2304 assert(verification_mark_stack()->isEmpty(), "markStack should be empty");
2305 verify_work_stacks_empty();
2306
2307 CMSHeap* heap = CMSHeap::heap();
2308 heap->ensure_parsability(false); // fill TLABs, but no need to retire them
2309 // Update the saved marks which may affect the root scans.
2310 heap->save_marks();
2311
2312 if (CMSRemarkVerifyVariant == 1) {
2313 // In this first variant of verification, we complete
2314 // all marking, then check if the new marks-vector is
2315 // a subset of the CMS marks-vector.
2316 verify_after_remark_work_1();
2317 } else {
2318 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2");
2319 // In this second variant of verification, we flag an error
2320 // (i.e. an object reachable in the new marks-vector not reachable
2321 // in the CMS marks-vector) immediately, also indicating the
2322 // identify of an object (A) that references the unmarked object (B) --
2323 // presumably, a mutation to A failed to be picked up by preclean/remark?
2324 verify_after_remark_work_2();
2325 }
2326
2327 return true;
2328 }
2329
2330 void CMSCollector::verify_after_remark_work_1() {
2331 ResourceMark rm;
2332 HandleMark hm;
2333 CMSHeap* heap = CMSHeap::heap();
2334
2335 // Get a clear set of claim bits for the roots processing to work with.
2336 ClassLoaderDataGraph::clear_claimed_marks();
2337
2338 // Mark from roots one level into CMS
2339 MarkRefsIntoClosure notOlder(_span, verification_mark_bm());
2340 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2341
2342 {
2343 StrongRootsScope srs(1);
2344
2345 heap->cms_process_roots(&srs,
2346 true, // young gen as roots
2347 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2348 should_unload_classes(),
2349 ¬Older,
2350 NULL);
2351 }
2352
2353 // Now mark from the roots
2354 MarkFromRootsClosure markFromRootsClosure(this, _span,
2355 verification_mark_bm(), verification_mark_stack(),
2356 false /* don't yield */, true /* verifying */);
2357 assert(_restart_addr == NULL, "Expected pre-condition");
2358 verification_mark_bm()->iterate(&markFromRootsClosure);
2359 while (_restart_addr != NULL) {
2360 // Deal with stack overflow: by restarting at the indicated
2361 // address.
2362 HeapWord* ra = _restart_addr;
2363 markFromRootsClosure.reset(ra);
2364 _restart_addr = NULL;
2365 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2366 }
2367 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2368 verify_work_stacks_empty();
2369
2370 // Marking completed -- now verify that each bit marked in
2371 // verification_mark_bm() is also marked in markBitMap(); flag all
2372 // errors by printing corresponding objects.
2373 VerifyMarkedClosure vcl(markBitMap());
2374 verification_mark_bm()->iterate(&vcl);
2375 if (vcl.failed()) {
2376 Log(gc, verify) log;
2377 log.error("Failed marking verification after remark");
2378 ResourceMark rm;
2379 LogStream ls(log.error());
2380 heap->print_on(&ls);
2381 fatal("CMS: failed marking verification after remark");
2382 }
2383 }
2384
2385 class VerifyCLDOopsCLDClosure : public CLDClosure {
2386 class VerifyCLDOopsClosure : public OopClosure {
2387 CMSBitMap* _bitmap;
2388 public:
2389 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { }
2390 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); }
2391 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
2392 } _oop_closure;
2393 public:
2394 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {}
2395 void do_cld(ClassLoaderData* cld) {
2396 cld->oops_do(&_oop_closure, false, false);
2397 }
2398 };
2399
2400 void CMSCollector::verify_after_remark_work_2() {
2401 ResourceMark rm;
2402 HandleMark hm;
2403 CMSHeap* heap = CMSHeap::heap();
2404
2405 // Get a clear set of claim bits for the roots processing to work with.
2406 ClassLoaderDataGraph::clear_claimed_marks();
2407
2408 // Mark from roots one level into CMS
2409 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(),
2410 markBitMap());
2411 CLDToOopClosure cld_closure(¬Older, true);
2412
2413 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2414
2415 {
2416 StrongRootsScope srs(1);
2417
2418 heap->cms_process_roots(&srs,
2419 true, // young gen as roots
2420 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2421 should_unload_classes(),
2422 ¬Older,
2423 &cld_closure);
2424 }
2425
2426 // Now mark from the roots
2427 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span,
2428 verification_mark_bm(), markBitMap(), verification_mark_stack());
2429 assert(_restart_addr == NULL, "Expected pre-condition");
2430 verification_mark_bm()->iterate(&markFromRootsClosure);
2431 while (_restart_addr != NULL) {
2432 // Deal with stack overflow: by restarting at the indicated
2433 // address.
2434 HeapWord* ra = _restart_addr;
2435 markFromRootsClosure.reset(ra);
2436 _restart_addr = NULL;
2437 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end());
2438 }
2439 assert(verification_mark_stack()->isEmpty(), "Should have been drained");
2440 verify_work_stacks_empty();
2441
2442 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm());
2443 ClassLoaderDataGraph::cld_do(&verify_cld_oops);
2444
2445 // Marking completed -- now verify that each bit marked in
2446 // verification_mark_bm() is also marked in markBitMap(); flag all
2447 // errors by printing corresponding objects.
2448 VerifyMarkedClosure vcl(markBitMap());
2449 verification_mark_bm()->iterate(&vcl);
2450 assert(!vcl.failed(), "Else verification above should not have succeeded");
2451 }
2452
2453 void ConcurrentMarkSweepGeneration::save_marks() {
2454 // delegate to CMS space
2455 cmsSpace()->save_marks();
2456 }
2457
2458 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() {
2459 return cmsSpace()->no_allocs_since_save_marks();
2460 }
2461
2462 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \
2463 \
2464 void ConcurrentMarkSweepGeneration:: \
2465 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \
2466 cl->set_generation(this); \
2467 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \
2468 cl->reset_generation(); \
2469 save_marks(); \
2470 }
2471
2472 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN)
2473
2474 void
2475 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) {
2476 if (freelistLock()->owned_by_self()) {
2477 Generation::oop_iterate(cl);
2478 } else {
2479 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2480 Generation::oop_iterate(cl);
2481 }
2482 }
2483
2484 void
2485 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) {
2486 if (freelistLock()->owned_by_self()) {
2487 Generation::object_iterate(cl);
2488 } else {
2489 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2490 Generation::object_iterate(cl);
2491 }
2492 }
2493
2494 void
2495 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) {
2496 if (freelistLock()->owned_by_self()) {
2497 Generation::safe_object_iterate(cl);
2498 } else {
2499 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2500 Generation::safe_object_iterate(cl);
2501 }
2502 }
2503
2504 void
2505 ConcurrentMarkSweepGeneration::post_compact() {
2506 }
2507
2508 void
2509 ConcurrentMarkSweepGeneration::prepare_for_verify() {
2510 // Fix the linear allocation blocks to look like free blocks.
2511
2512 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2513 // are not called when the heap is verified during universe initialization and
2514 // at vm shutdown.
2515 if (freelistLock()->owned_by_self()) {
2516 cmsSpace()->prepare_for_verify();
2517 } else {
2518 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2519 cmsSpace()->prepare_for_verify();
2520 }
2521 }
2522
2523 void
2524 ConcurrentMarkSweepGeneration::verify() {
2525 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those
2526 // are not called when the heap is verified during universe initialization and
2527 // at vm shutdown.
2528 if (freelistLock()->owned_by_self()) {
2529 cmsSpace()->verify();
2530 } else {
2531 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag);
2532 cmsSpace()->verify();
2533 }
2534 }
2535
2536 void CMSCollector::verify() {
2537 _cmsGen->verify();
2538 }
2539
2540 #ifndef PRODUCT
2541 bool CMSCollector::overflow_list_is_empty() const {
2542 assert(_num_par_pushes >= 0, "Inconsistency");
2543 if (_overflow_list == NULL) {
2544 assert(_num_par_pushes == 0, "Inconsistency");
2545 }
2546 return _overflow_list == NULL;
2547 }
2548
2549 // The methods verify_work_stacks_empty() and verify_overflow_empty()
2550 // merely consolidate assertion checks that appear to occur together frequently.
2551 void CMSCollector::verify_work_stacks_empty() const {
2552 assert(_markStack.isEmpty(), "Marking stack should be empty");
2553 assert(overflow_list_is_empty(), "Overflow list should be empty");
2554 }
2555
2556 void CMSCollector::verify_overflow_empty() const {
2557 assert(overflow_list_is_empty(), "Overflow list should be empty");
2558 assert(no_preserved_marks(), "No preserved marks");
2559 }
2560 #endif // PRODUCT
2561
2562 // Decide if we want to enable class unloading as part of the
2563 // ensuing concurrent GC cycle. We will collect and
2564 // unload classes if it's the case that:
2565 // (a) class unloading is enabled at the command line, and
2566 // (b) old gen is getting really full
2567 // NOTE: Provided there is no change in the state of the heap between
2568 // calls to this method, it should have idempotent results. Moreover,
2569 // its results should be monotonically increasing (i.e. going from 0 to 1,
2570 // but not 1 to 0) between successive calls between which the heap was
2571 // not collected. For the implementation below, it must thus rely on
2572 // the property that concurrent_cycles_since_last_unload()
2573 // will not decrease unless a collection cycle happened and that
2574 // _cmsGen->is_too_full() are
2575 // themselves also monotonic in that sense. See check_monotonicity()
2576 // below.
2577 void CMSCollector::update_should_unload_classes() {
2578 _should_unload_classes = false;
2579 if (CMSClassUnloadingEnabled) {
2580 _should_unload_classes = (concurrent_cycles_since_last_unload() >=
2581 CMSClassUnloadingMaxInterval)
2582 || _cmsGen->is_too_full();
2583 }
2584 }
2585
2586 bool ConcurrentMarkSweepGeneration::is_too_full() const {
2587 bool res = should_concurrent_collect();
2588 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0);
2589 return res;
2590 }
2591
2592 void CMSCollector::setup_cms_unloading_and_verification_state() {
2593 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC
2594 || VerifyBeforeExit;
2595 const int rso = GenCollectedHeap::SO_AllCodeCache;
2596
2597 // We set the proper root for this CMS cycle here.
2598 if (should_unload_classes()) { // Should unload classes this cycle
2599 remove_root_scanning_option(rso); // Shrink the root set appropriately
2600 set_verifying(should_verify); // Set verification state for this cycle
2601 return; // Nothing else needs to be done at this time
2602 }
2603
2604 // Not unloading classes this cycle
2605 assert(!should_unload_classes(), "Inconsistency!");
2606
2607 // If we are not unloading classes then add SO_AllCodeCache to root
2608 // scanning options.
2609 add_root_scanning_option(rso);
2610
2611 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) {
2612 set_verifying(true);
2613 } else if (verifying() && !should_verify) {
2614 // We were verifying, but some verification flags got disabled.
2615 set_verifying(false);
2616 // Exclude symbols, strings and code cache elements from root scanning to
2617 // reduce IM and RM pauses.
2618 remove_root_scanning_option(rso);
2619 }
2620 }
2621
2622
2623 #ifndef PRODUCT
2624 HeapWord* CMSCollector::block_start(const void* p) const {
2625 const HeapWord* addr = (HeapWord*)p;
2626 if (_span.contains(p)) {
2627 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) {
2628 return _cmsGen->cmsSpace()->block_start(p);
2629 }
2630 }
2631 return NULL;
2632 }
2633 #endif
2634
2635 HeapWord*
2636 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
2637 bool tlab,
2638 bool parallel) {
2639 CMSSynchronousYieldRequest yr;
2640 assert(!tlab, "Can't deal with TLAB allocation");
2641 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
2642 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation);
2643 if (GCExpandToAllocateDelayMillis > 0) {
2644 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2645 }
2646 return have_lock_and_allocate(word_size, tlab);
2647 }
2648
2649 void ConcurrentMarkSweepGeneration::expand_for_gc_cause(
2650 size_t bytes,
2651 size_t expand_bytes,
2652 CMSExpansionCause::Cause cause)
2653 {
2654
2655 bool success = expand(bytes, expand_bytes);
2656
2657 // remember why we expanded; this information is used
2658 // by shouldConcurrentCollect() when making decisions on whether to start
2659 // a new CMS cycle.
2660 if (success) {
2661 set_expansion_cause(cause);
2662 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause));
2663 }
2664 }
2665
2666 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) {
2667 HeapWord* res = NULL;
2668 MutexLocker x(ParGCRareEvent_lock);
2669 while (true) {
2670 // Expansion by some other thread might make alloc OK now:
2671 res = ps->lab.alloc(word_sz);
2672 if (res != NULL) return res;
2673 // If there's not enough expansion space available, give up.
2674 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) {
2675 return NULL;
2676 }
2677 // Otherwise, we try expansion.
2678 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab);
2679 // Now go around the loop and try alloc again;
2680 // A competing par_promote might beat us to the expansion space,
2681 // so we may go around the loop again if promotion fails again.
2682 if (GCExpandToAllocateDelayMillis > 0) {
2683 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2684 }
2685 }
2686 }
2687
2688
2689 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space(
2690 PromotionInfo* promo) {
2691 MutexLocker x(ParGCRareEvent_lock);
2692 size_t refill_size_bytes = promo->refillSize() * HeapWordSize;
2693 while (true) {
2694 // Expansion by some other thread might make alloc OK now:
2695 if (promo->ensure_spooling_space()) {
2696 assert(promo->has_spooling_space(),
2697 "Post-condition of successful ensure_spooling_space()");
2698 return true;
2699 }
2700 // If there's not enough expansion space available, give up.
2701 if (_virtual_space.uncommitted_size() < refill_size_bytes) {
2702 return false;
2703 }
2704 // Otherwise, we try expansion.
2705 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space);
2706 // Now go around the loop and try alloc again;
2707 // A competing allocation might beat us to the expansion space,
2708 // so we may go around the loop again if allocation fails again.
2709 if (GCExpandToAllocateDelayMillis > 0) {
2710 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false);
2711 }
2712 }
2713 }
2714
2715 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) {
2716 // Only shrink if a compaction was done so that all the free space
2717 // in the generation is in a contiguous block at the end.
2718 if (did_compact()) {
2719 CardGeneration::shrink(bytes);
2720 }
2721 }
2722
2723 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() {
2724 assert_locked_or_safepoint(Heap_lock);
2725 }
2726
2727 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) {
2728 assert_locked_or_safepoint(Heap_lock);
2729 assert_lock_strong(freelistLock());
2730 log_trace(gc)("Shrinking of CMS not yet implemented");
2731 return;
2732 }
2733
2734
2735 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent
2736 // phases.
2737 class CMSPhaseAccounting: public StackObj {
2738 public:
2739 CMSPhaseAccounting(CMSCollector *collector,
2740 const char *title);
2741 ~CMSPhaseAccounting();
2742
2743 private:
2744 CMSCollector *_collector;
2745 const char *_title;
2746 GCTraceConcTime(Info, gc) _trace_time;
2747
2748 public:
2749 // Not MT-safe; so do not pass around these StackObj's
2750 // where they may be accessed by other threads.
2751 double wallclock_millis() {
2752 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time());
2753 }
2754 };
2755
2756 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector,
2757 const char *title) :
2758 _collector(collector), _title(title), _trace_time(title) {
2759
2760 _collector->resetYields();
2761 _collector->resetTimer();
2762 _collector->startTimer();
2763 _collector->gc_timer_cm()->register_gc_concurrent_start(title);
2764 }
2765
2766 CMSPhaseAccounting::~CMSPhaseAccounting() {
2767 _collector->gc_timer_cm()->register_gc_concurrent_end();
2768 _collector->stopTimer();
2769 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks()));
2770 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields());
2771 }
2772
2773 // CMS work
2774
2775 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask.
2776 class CMSParMarkTask : public AbstractGangTask {
2777 protected:
2778 CMSCollector* _collector;
2779 uint _n_workers;
2780 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) :
2781 AbstractGangTask(name),
2782 _collector(collector),
2783 _n_workers(n_workers) {}
2784 // Work method in support of parallel rescan ... of young gen spaces
2785 void do_young_space_rescan(OopsInGenClosure* cl,
2786 ContiguousSpace* space,
2787 HeapWord** chunk_array, size_t chunk_top);
2788 void work_on_young_gen_roots(OopsInGenClosure* cl);
2789 };
2790
2791 // Parallel initial mark task
2792 class CMSParInitialMarkTask: public CMSParMarkTask {
2793 StrongRootsScope* _strong_roots_scope;
2794 public:
2795 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) :
2796 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers),
2797 _strong_roots_scope(strong_roots_scope) {}
2798 void work(uint worker_id);
2799 };
2800
2801 // Checkpoint the roots into this generation from outside
2802 // this generation. [Note this initial checkpoint need only
2803 // be approximate -- we'll do a catch up phase subsequently.]
2804 void CMSCollector::checkpointRootsInitial() {
2805 assert(_collectorState == InitialMarking, "Wrong collector state");
2806 check_correct_thread_executing();
2807 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
2808
2809 save_heap_summary();
2810 report_heap_summary(GCWhen::BeforeGC);
2811
2812 ReferenceProcessor* rp = ref_processor();
2813 assert(_restart_addr == NULL, "Control point invariant");
2814 {
2815 // acquire locks for subsequent manipulations
2816 MutexLockerEx x(bitMapLock(),
2817 Mutex::_no_safepoint_check_flag);
2818 checkpointRootsInitialWork();
2819 // enable ("weak") refs discovery
2820 rp->enable_discovery();
2821 _collectorState = Marking;
2822 }
2823 }
2824
2825 void CMSCollector::checkpointRootsInitialWork() {
2826 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
2827 assert(_collectorState == InitialMarking, "just checking");
2828
2829 // Already have locks.
2830 assert_lock_strong(bitMapLock());
2831 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle");
2832
2833 // Setup the verification and class unloading state for this
2834 // CMS collection cycle.
2835 setup_cms_unloading_and_verification_state();
2836
2837 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm);
2838
2839 // Reset all the PLAB chunk arrays if necessary.
2840 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) {
2841 reset_survivor_plab_arrays();
2842 }
2843
2844 ResourceMark rm;
2845 HandleMark hm;
2846
2847 MarkRefsIntoClosure notOlder(_span, &_markBitMap);
2848 CMSHeap* heap = CMSHeap::heap();
2849
2850 verify_work_stacks_empty();
2851 verify_overflow_empty();
2852
2853 heap->ensure_parsability(false); // fill TLABs, but no need to retire them
2854 // Update the saved marks which may affect the root scans.
2855 heap->save_marks();
2856
2857 // weak reference processing has not started yet.
2858 ref_processor()->set_enqueuing_is_done(false);
2859
2860 // Need to remember all newly created CLDs,
2861 // so that we can guarantee that the remark finds them.
2862 ClassLoaderDataGraph::remember_new_clds(true);
2863
2864 // Whenever a CLD is found, it will be claimed before proceeding to mark
2865 // the klasses. The claimed marks need to be cleared before marking starts.
2866 ClassLoaderDataGraph::clear_claimed_marks();
2867
2868 print_eden_and_survivor_chunk_arrays();
2869
2870 {
2871 #if COMPILER2_OR_JVMCI
2872 DerivedPointerTableDeactivate dpt_deact;
2873 #endif
2874 if (CMSParallelInitialMarkEnabled) {
2875 // The parallel version.
2876 WorkGang* workers = heap->workers();
2877 assert(workers != NULL, "Need parallel worker threads.");
2878 uint n_workers = workers->active_workers();
2879
2880 StrongRootsScope srs(n_workers);
2881
2882 CMSParInitialMarkTask tsk(this, &srs, n_workers);
2883 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
2884 // If the total workers is greater than 1, then multiple workers
2885 // may be used at some time and the initialization has been set
2886 // such that the single threaded path cannot be used.
2887 if (workers->total_workers() > 1) {
2888 workers->run_task(&tsk);
2889 } else {
2890 tsk.work(0);
2891 }
2892 } else {
2893 // The serial version.
2894 CLDToOopClosure cld_closure(¬Older, true);
2895 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
2896
2897 StrongRootsScope srs(1);
2898
2899 heap->cms_process_roots(&srs,
2900 true, // young gen as roots
2901 GenCollectedHeap::ScanningOption(roots_scanning_options()),
2902 should_unload_classes(),
2903 ¬Older,
2904 &cld_closure);
2905 }
2906 }
2907
2908 // Clear mod-union table; it will be dirtied in the prologue of
2909 // CMS generation per each young generation collection.
2910
2911 assert(_modUnionTable.isAllClear(),
2912 "Was cleared in most recent final checkpoint phase"
2913 " or no bits are set in the gc_prologue before the start of the next "
2914 "subsequent marking phase.");
2915
2916 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be");
2917
2918 // Save the end of the used_region of the constituent generations
2919 // to be used to limit the extent of sweep in each generation.
2920 save_sweep_limits();
2921 verify_overflow_empty();
2922 }
2923
2924 bool CMSCollector::markFromRoots() {
2925 // we might be tempted to assert that:
2926 // assert(!SafepointSynchronize::is_at_safepoint(),
2927 // "inconsistent argument?");
2928 // However that wouldn't be right, because it's possible that
2929 // a safepoint is indeed in progress as a young generation
2930 // stop-the-world GC happens even as we mark in this generation.
2931 assert(_collectorState == Marking, "inconsistent state?");
2932 check_correct_thread_executing();
2933 verify_overflow_empty();
2934
2935 // Weak ref discovery note: We may be discovering weak
2936 // refs in this generation concurrent (but interleaved) with
2937 // weak ref discovery by the young generation collector.
2938
2939 CMSTokenSyncWithLocks ts(true, bitMapLock());
2940 GCTraceCPUTime tcpu;
2941 CMSPhaseAccounting pa(this, "Concurrent Mark");
2942 bool res = markFromRootsWork();
2943 if (res) {
2944 _collectorState = Precleaning;
2945 } else { // We failed and a foreground collection wants to take over
2946 assert(_foregroundGCIsActive, "internal state inconsistency");
2947 assert(_restart_addr == NULL, "foreground will restart from scratch");
2948 log_debug(gc)("bailing out to foreground collection");
2949 }
2950 verify_overflow_empty();
2951 return res;
2952 }
2953
2954 bool CMSCollector::markFromRootsWork() {
2955 // iterate over marked bits in bit map, doing a full scan and mark
2956 // from these roots using the following algorithm:
2957 // . if oop is to the right of the current scan pointer,
2958 // mark corresponding bit (we'll process it later)
2959 // . else (oop is to left of current scan pointer)
2960 // push oop on marking stack
2961 // . drain the marking stack
2962
2963 // Note that when we do a marking step we need to hold the
2964 // bit map lock -- recall that direct allocation (by mutators)
2965 // and promotion (by the young generation collector) is also
2966 // marking the bit map. [the so-called allocate live policy.]
2967 // Because the implementation of bit map marking is not
2968 // robust wrt simultaneous marking of bits in the same word,
2969 // we need to make sure that there is no such interference
2970 // between concurrent such updates.
2971
2972 // already have locks
2973 assert_lock_strong(bitMapLock());
2974
2975 verify_work_stacks_empty();
2976 verify_overflow_empty();
2977 bool result = false;
2978 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) {
2979 result = do_marking_mt();
2980 } else {
2981 result = do_marking_st();
2982 }
2983 return result;
2984 }
2985
2986 // Forward decl
2987 class CMSConcMarkingTask;
2988
2989 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
2990 CMSCollector* _collector;
2991 CMSConcMarkingTask* _task;
2992 public:
2993 virtual void yield();
2994
2995 // "n_threads" is the number of threads to be terminated.
2996 // "queue_set" is a set of work queues of other threads.
2997 // "collector" is the CMS collector associated with this task terminator.
2998 // "yield" indicates whether we need the gang as a whole to yield.
2999 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
3000 ParallelTaskTerminator(n_threads, queue_set),
3001 _collector(collector) { }
3002
3003 void set_task(CMSConcMarkingTask* task) {
3004 _task = task;
3005 }
3006 };
3007
3008 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
3009 CMSConcMarkingTask* _task;
3010 public:
3011 bool should_exit_termination();
3012 void set_task(CMSConcMarkingTask* task) {
3013 _task = task;
3014 }
3015 };
3016
3017 // MT Concurrent Marking Task
3018 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
3019 CMSCollector* _collector;
3020 uint _n_workers; // requested/desired # workers
3021 bool _result;
3022 CompactibleFreeListSpace* _cms_space;
3023 char _pad_front[64]; // padding to ...
3024 HeapWord* volatile _global_finger; // ... avoid sharing cache line
3025 char _pad_back[64];
3026 HeapWord* _restart_addr;
3027
3028 // Exposed here for yielding support
3029 Mutex* const _bit_map_lock;
3030
3031 // The per thread work queues, available here for stealing
3032 OopTaskQueueSet* _task_queues;
3033
3034 // Termination (and yielding) support
3035 CMSConcMarkingTerminator _term;
3036 CMSConcMarkingTerminatorTerminator _term_term;
3037
3038 public:
3039 CMSConcMarkingTask(CMSCollector* collector,
3040 CompactibleFreeListSpace* cms_space,
3041 YieldingFlexibleWorkGang* workers,
3042 OopTaskQueueSet* task_queues):
3043 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
3044 _collector(collector),
3045 _cms_space(cms_space),
3046 _n_workers(0), _result(true),
3047 _task_queues(task_queues),
3048 _term(_n_workers, task_queues, _collector),
3049 _bit_map_lock(collector->bitMapLock())
3050 {
3051 _requested_size = _n_workers;
3052 _term.set_task(this);
3053 _term_term.set_task(this);
3054 _restart_addr = _global_finger = _cms_space->bottom();
3055 }
3056
3057
3058 OopTaskQueueSet* task_queues() { return _task_queues; }
3059
3060 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
3061
3062 HeapWord* volatile* global_finger_addr() { return &_global_finger; }
3063
3064 CMSConcMarkingTerminator* terminator() { return &_term; }
3065
3066 virtual void set_for_termination(uint active_workers) {
3067 terminator()->reset_for_reuse(active_workers);
3068 }
3069
3070 void work(uint worker_id);
3071 bool should_yield() {
3072 return ConcurrentMarkSweepThread::should_yield()
3073 && !_collector->foregroundGCIsActive();
3074 }
3075
3076 virtual void coordinator_yield(); // stuff done by coordinator
3077 bool result() { return _result; }
3078
3079 void reset(HeapWord* ra) {
3080 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)");
3081 _restart_addr = _global_finger = ra;
3082 _term.reset_for_reuse();
3083 }
3084
3085 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3086 OopTaskQueue* work_q);
3087
3088 private:
3089 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp);
3090 void do_work_steal(int i);
3091 void bump_global_finger(HeapWord* f);
3092 };
3093
3094 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
3095 assert(_task != NULL, "Error");
3096 return _task->yielding();
3097 // Note that we do not need the disjunct || _task->should_yield() above
3098 // because we want terminating threads to yield only if the task
3099 // is already in the midst of yielding, which happens only after at least one
3100 // thread has yielded.
3101 }
3102
3103 void CMSConcMarkingTerminator::yield() {
3104 if (_task->should_yield()) {
3105 _task->yield();
3106 } else {
3107 ParallelTaskTerminator::yield();
3108 }
3109 }
3110
3111 ////////////////////////////////////////////////////////////////
3112 // Concurrent Marking Algorithm Sketch
3113 ////////////////////////////////////////////////////////////////
3114 // Until all tasks exhausted (both spaces):
3115 // -- claim next available chunk
3116 // -- bump global finger via CAS
3117 // -- find first object that starts in this chunk
3118 // and start scanning bitmap from that position
3119 // -- scan marked objects for oops
3120 // -- CAS-mark target, and if successful:
3121 // . if target oop is above global finger (volatile read)
3122 // nothing to do
3123 // . if target oop is in chunk and above local finger
3124 // then nothing to do
3125 // . else push on work-queue
3126 // -- Deal with possible overflow issues:
3127 // . local work-queue overflow causes stuff to be pushed on
3128 // global (common) overflow queue
3129 // . always first empty local work queue
3130 // . then get a batch of oops from global work queue if any
3131 // . then do work stealing
3132 // -- When all tasks claimed (both spaces)
3133 // and local work queue empty,
3134 // then in a loop do:
3135 // . check global overflow stack; steal a batch of oops and trace
3136 // . try to steal from other threads oif GOS is empty
3137 // . if neither is available, offer termination
3138 // -- Terminate and return result
3139 //
3140 void CMSConcMarkingTask::work(uint worker_id) {
3141 elapsedTimer _timer;
3142 ResourceMark rm;
3143 HandleMark hm;
3144
3145 DEBUG_ONLY(_collector->verify_overflow_empty();)
3146
3147 // Before we begin work, our work queue should be empty
3148 assert(work_queue(worker_id)->size() == 0, "Expected to be empty");
3149 // Scan the bitmap covering _cms_space, tracing through grey objects.
3150 _timer.start();
3151 do_scan_and_mark(worker_id, _cms_space);
3152 _timer.stop();
3153 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3154
3155 // ... do work stealing
3156 _timer.reset();
3157 _timer.start();
3158 do_work_steal(worker_id);
3159 _timer.stop();
3160 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
3161 assert(_collector->_markStack.isEmpty(), "Should have been emptied");
3162 assert(work_queue(worker_id)->size() == 0, "Should have been emptied");
3163 // Note that under the current task protocol, the
3164 // following assertion is true even of the spaces
3165 // expanded since the completion of the concurrent
3166 // marking. XXX This will likely change under a strict
3167 // ABORT semantics.
3168 // After perm removal the comparison was changed to
3169 // greater than or equal to from strictly greater than.
3170 // Before perm removal the highest address sweep would
3171 // have been at the end of perm gen but now is at the
3172 // end of the tenured gen.
3173 assert(_global_finger >= _cms_space->end(),
3174 "All tasks have been completed");
3175 DEBUG_ONLY(_collector->verify_overflow_empty();)
3176 }
3177
3178 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) {
3179 HeapWord* read = _global_finger;
3180 HeapWord* cur = read;
3181 while (f > read) {
3182 cur = read;
3183 read = Atomic::cmpxchg(f, &_global_finger, cur);
3184 if (cur == read) {
3185 // our cas succeeded
3186 assert(_global_finger >= f, "protocol consistency");
3187 break;
3188 }
3189 }
3190 }
3191
3192 // This is really inefficient, and should be redone by
3193 // using (not yet available) block-read and -write interfaces to the
3194 // stack and the work_queue. XXX FIX ME !!!
3195 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk,
3196 OopTaskQueue* work_q) {
3197 // Fast lock-free check
3198 if (ovflw_stk->length() == 0) {
3199 return false;
3200 }
3201 assert(work_q->size() == 0, "Shouldn't steal");
3202 MutexLockerEx ml(ovflw_stk->par_lock(),
3203 Mutex::_no_safepoint_check_flag);
3204 // Grab up to 1/4 the size of the work queue
3205 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
3206 (size_t)ParGCDesiredObjsFromOverflowList);
3207 num = MIN2(num, ovflw_stk->length());
3208 for (int i = (int) num; i > 0; i--) {
3209 oop cur = ovflw_stk->pop();
3210 assert(cur != NULL, "Counted wrong?");
3211 work_q->push(cur);
3212 }
3213 return num > 0;
3214 }
3215
3216 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) {
3217 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
3218 int n_tasks = pst->n_tasks();
3219 // We allow that there may be no tasks to do here because
3220 // we are restarting after a stack overflow.
3221 assert(pst->valid() || n_tasks == 0, "Uninitialized use?");
3222 uint nth_task = 0;
3223
3224 HeapWord* aligned_start = sp->bottom();
3225 if (sp->used_region().contains(_restart_addr)) {
3226 // Align down to a card boundary for the start of 0th task
3227 // for this space.
3228 aligned_start = align_down(_restart_addr, CardTable::card_size);
3229 }
3230
3231 size_t chunk_size = sp->marking_task_size();
3232 while (!pst->is_task_claimed(/* reference */ nth_task)) {
3233 // Having claimed the nth task in this space,
3234 // compute the chunk that it corresponds to:
3235 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size,
3236 aligned_start + (nth_task+1)*chunk_size);
3237 // Try and bump the global finger via a CAS;
3238 // note that we need to do the global finger bump
3239 // _before_ taking the intersection below, because
3240 // the task corresponding to that region will be
3241 // deemed done even if the used_region() expands
3242 // because of allocation -- as it almost certainly will
3243 // during start-up while the threads yield in the
3244 // closure below.
3245 HeapWord* finger = span.end();
3246 bump_global_finger(finger); // atomically
3247 // There are null tasks here corresponding to chunks
3248 // beyond the "top" address of the space.
3249 span = span.intersection(sp->used_region());
3250 if (!span.is_empty()) { // Non-null task
3251 HeapWord* prev_obj;
3252 assert(!span.contains(_restart_addr) || nth_task == 0,
3253 "Inconsistency");
3254 if (nth_task == 0) {
3255 // For the 0th task, we'll not need to compute a block_start.
3256 if (span.contains(_restart_addr)) {
3257 // In the case of a restart because of stack overflow,
3258 // we might additionally skip a chunk prefix.
3259 prev_obj = _restart_addr;
3260 } else {
3261 prev_obj = span.start();
3262 }
3263 } else {
3264 // We want to skip the first object because
3265 // the protocol is to scan any object in its entirety
3266 // that _starts_ in this span; a fortiori, any
3267 // object starting in an earlier span is scanned
3268 // as part of an earlier claimed task.
3269 // Below we use the "careful" version of block_start
3270 // so we do not try to navigate uninitialized objects.
3271 prev_obj = sp->block_start_careful(span.start());
3272 // Below we use a variant of block_size that uses the
3273 // Printezis bits to avoid waiting for allocated
3274 // objects to become initialized/parsable.
3275 while (prev_obj < span.start()) {
3276 size_t sz = sp->block_size_no_stall(prev_obj, _collector);
3277 if (sz > 0) {
3278 prev_obj += sz;
3279 } else {
3280 // In this case we may end up doing a bit of redundant
3281 // scanning, but that appears unavoidable, short of
3282 // locking the free list locks; see bug 6324141.
3283 break;
3284 }
3285 }
3286 }
3287 if (prev_obj < span.end()) {
3288 MemRegion my_span = MemRegion(prev_obj, span.end());
3289 // Do the marking work within a non-empty span --
3290 // the last argument to the constructor indicates whether the
3291 // iteration should be incremental with periodic yields.
3292 ParMarkFromRootsClosure cl(this, _collector, my_span,
3293 &_collector->_markBitMap,
3294 work_queue(i),
3295 &_collector->_markStack);
3296 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end());
3297 } // else nothing to do for this task
3298 } // else nothing to do for this task
3299 }
3300 // We'd be tempted to assert here that since there are no
3301 // more tasks left to claim in this space, the global_finger
3302 // must exceed space->top() and a fortiori space->end(). However,
3303 // that would not quite be correct because the bumping of
3304 // global_finger occurs strictly after the claiming of a task,
3305 // so by the time we reach here the global finger may not yet
3306 // have been bumped up by the thread that claimed the last
3307 // task.
3308 pst->all_tasks_completed();
3309 }
3310
3311 class ParConcMarkingClosure: public MetadataAwareOopClosure {
3312 private:
3313 CMSCollector* _collector;
3314 CMSConcMarkingTask* _task;
3315 MemRegion _span;
3316 CMSBitMap* _bit_map;
3317 CMSMarkStack* _overflow_stack;
3318 OopTaskQueue* _work_queue;
3319 protected:
3320 DO_OOP_WORK_DEFN
3321 public:
3322 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
3323 CMSBitMap* bit_map, CMSMarkStack* overflow_stack):
3324 MetadataAwareOopClosure(collector->ref_processor()),
3325 _collector(collector),
3326 _task(task),
3327 _span(collector->_span),
3328 _work_queue(work_queue),
3329 _bit_map(bit_map),
3330 _overflow_stack(overflow_stack)
3331 { }
3332 virtual void do_oop(oop* p);
3333 virtual void do_oop(narrowOop* p);
3334
3335 void trim_queue(size_t max);
3336 void handle_stack_overflow(HeapWord* lost);
3337 void do_yield_check() {
3338 if (_task->should_yield()) {
3339 _task->yield();
3340 }
3341 }
3342 };
3343
3344 DO_OOP_WORK_IMPL(ParConcMarkingClosure)
3345
3346 // Grey object scanning during work stealing phase --
3347 // the salient assumption here is that any references
3348 // that are in these stolen objects being scanned must
3349 // already have been initialized (else they would not have
3350 // been published), so we do not need to check for
3351 // uninitialized objects before pushing here.
3352 void ParConcMarkingClosure::do_oop(oop obj) {
3353 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
3354 HeapWord* addr = (HeapWord*)obj;
3355 // Check if oop points into the CMS generation
3356 // and is not marked
3357 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
3358 // a white object ...
3359 // If we manage to "claim" the object, by being the
3360 // first thread to mark it, then we push it on our
3361 // marking stack
3362 if (_bit_map->par_mark(addr)) { // ... now grey
3363 // push on work queue (grey set)
3364 bool simulate_overflow = false;
3365 NOT_PRODUCT(
3366 if (CMSMarkStackOverflowALot &&
3367 _collector->simulate_overflow()) {
3368 // simulate a stack overflow
3369 simulate_overflow = true;
3370 }
3371 )
3372 if (simulate_overflow ||
3373 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
3374 // stack overflow
3375 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
3376 // We cannot assert that the overflow stack is full because
3377 // it may have been emptied since.
3378 assert(simulate_overflow ||
3379 _work_queue->size() == _work_queue->max_elems(),
3380 "Else push should have succeeded");
3381 handle_stack_overflow(addr);
3382 }
3383 } // Else, some other thread got there first
3384 do_yield_check();
3385 }
3386 }
3387
3388 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); }
3389 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); }
3390
3391 void ParConcMarkingClosure::trim_queue(size_t max) {
3392 while (_work_queue->size() > max) {
3393 oop new_oop;
3394 if (_work_queue->pop_local(new_oop)) {
3395 assert(oopDesc::is_oop(new_oop), "Should be an oop");
3396 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object");
3397 assert(_span.contains((HeapWord*)new_oop), "Not in span");
3398 new_oop->oop_iterate(this); // do_oop() above
3399 do_yield_check();
3400 }
3401 }
3402 }
3403
3404 // Upon stack overflow, we discard (part of) the stack,
3405 // remembering the least address amongst those discarded
3406 // in CMSCollector's _restart_address.
3407 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) {
3408 // We need to do this under a mutex to prevent other
3409 // workers from interfering with the work done below.
3410 MutexLockerEx ml(_overflow_stack->par_lock(),
3411 Mutex::_no_safepoint_check_flag);
3412 // Remember the least grey address discarded
3413 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
3414 _collector->lower_restart_addr(ra);
3415 _overflow_stack->reset(); // discard stack contents
3416 _overflow_stack->expand(); // expand the stack if possible
3417 }
3418
3419
3420 void CMSConcMarkingTask::do_work_steal(int i) {
3421 OopTaskQueue* work_q = work_queue(i);
3422 oop obj_to_scan;
3423 CMSBitMap* bm = &(_collector->_markBitMap);
3424 CMSMarkStack* ovflw = &(_collector->_markStack);
3425 int* seed = _collector->hash_seed(i);
3426 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
3427 while (true) {
3428 cl.trim_queue(0);
3429 assert(work_q->size() == 0, "Should have been emptied above");
3430 if (get_work_from_overflow_stack(ovflw, work_q)) {
3431 // Can't assert below because the work obtained from the
3432 // overflow stack may already have been stolen from us.
3433 // assert(work_q->size() > 0, "Work from overflow stack");
3434 continue;
3435 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
3436 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop");
3437 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
3438 obj_to_scan->oop_iterate(&cl);
3439 } else if (terminator()->offer_termination(&_term_term)) {
3440 assert(work_q->size() == 0, "Impossible!");
3441 break;
3442 } else if (yielding() || should_yield()) {
3443 yield();
3444 }
3445 }
3446 }
3447
3448 // This is run by the CMS (coordinator) thread.
3449 void CMSConcMarkingTask::coordinator_yield() {
3450 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3451 "CMS thread should hold CMS token");
3452 // First give up the locks, then yield, then re-lock
3453 // We should probably use a constructor/destructor idiom to
3454 // do this unlock/lock or modify the MutexUnlocker class to
3455 // serve our purpose. XXX
3456 assert_lock_strong(_bit_map_lock);
3457 _bit_map_lock->unlock();
3458 ConcurrentMarkSweepThread::desynchronize(true);
3459 _collector->stopTimer();
3460 _collector->incrementYields();
3461
3462 // It is possible for whichever thread initiated the yield request
3463 // not to get a chance to wake up and take the bitmap lock between
3464 // this thread releasing it and reacquiring it. So, while the
3465 // should_yield() flag is on, let's sleep for a bit to give the
3466 // other thread a chance to wake up. The limit imposed on the number
3467 // of iterations is defensive, to avoid any unforseen circumstances
3468 // putting us into an infinite loop. Since it's always been this
3469 // (coordinator_yield()) method that was observed to cause the
3470 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount)
3471 // which is by default non-zero. For the other seven methods that
3472 // also perform the yield operation, as are using a different
3473 // parameter (CMSYieldSleepCount) which is by default zero. This way we
3474 // can enable the sleeping for those methods too, if necessary.
3475 // See 6442774.
3476 //
3477 // We really need to reconsider the synchronization between the GC
3478 // thread and the yield-requesting threads in the future and we
3479 // should really use wait/notify, which is the recommended
3480 // way of doing this type of interaction. Additionally, we should
3481 // consolidate the eight methods that do the yield operation and they
3482 // are almost identical into one for better maintainability and
3483 // readability. See 6445193.
3484 //
3485 // Tony 2006.06.29
3486 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount &&
3487 ConcurrentMarkSweepThread::should_yield() &&
3488 !CMSCollector::foregroundGCIsActive(); ++i) {
3489 os::sleep(Thread::current(), 1, false);
3490 }
3491
3492 ConcurrentMarkSweepThread::synchronize(true);
3493 _bit_map_lock->lock_without_safepoint_check();
3494 _collector->startTimer();
3495 }
3496
3497 bool CMSCollector::do_marking_mt() {
3498 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition");
3499 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(),
3500 conc_workers()->active_workers(),
3501 Threads::number_of_non_daemon_threads());
3502 num_workers = conc_workers()->update_active_workers(num_workers);
3503 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers());
3504
3505 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
3506
3507 CMSConcMarkingTask tsk(this,
3508 cms_space,
3509 conc_workers(),
3510 task_queues());
3511
3512 // Since the actual number of workers we get may be different
3513 // from the number we requested above, do we need to do anything different
3514 // below? In particular, may be we need to subclass the SequantialSubTasksDone
3515 // class?? XXX
3516 cms_space ->initialize_sequential_subtasks_for_marking(num_workers);
3517
3518 // Refs discovery is already non-atomic.
3519 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic");
3520 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT");
3521 conc_workers()->start_task(&tsk);
3522 while (tsk.yielded()) {
3523 tsk.coordinator_yield();
3524 conc_workers()->continue_task(&tsk);
3525 }
3526 // If the task was aborted, _restart_addr will be non-NULL
3527 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency");
3528 while (_restart_addr != NULL) {
3529 // XXX For now we do not make use of ABORTED state and have not
3530 // yet implemented the right abort semantics (even in the original
3531 // single-threaded CMS case). That needs some more investigation
3532 // and is deferred for now; see CR# TBF. 07252005YSR. XXX
3533 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency");
3534 // If _restart_addr is non-NULL, a marking stack overflow
3535 // occurred; we need to do a fresh marking iteration from the
3536 // indicated restart address.
3537 if (_foregroundGCIsActive) {
3538 // We may be running into repeated stack overflows, having
3539 // reached the limit of the stack size, while making very
3540 // slow forward progress. It may be best to bail out and
3541 // let the foreground collector do its job.
3542 // Clear _restart_addr, so that foreground GC
3543 // works from scratch. This avoids the headache of
3544 // a "rescan" which would otherwise be needed because
3545 // of the dirty mod union table & card table.
3546 _restart_addr = NULL;
3547 return false;
3548 }
3549 // Adjust the task to restart from _restart_addr
3550 tsk.reset(_restart_addr);
3551 cms_space ->initialize_sequential_subtasks_for_marking(num_workers,
3552 _restart_addr);
3553 _restart_addr = NULL;
3554 // Get the workers going again
3555 conc_workers()->start_task(&tsk);
3556 while (tsk.yielded()) {
3557 tsk.coordinator_yield();
3558 conc_workers()->continue_task(&tsk);
3559 }
3560 }
3561 assert(tsk.completed(), "Inconsistency");
3562 assert(tsk.result() == true, "Inconsistency");
3563 return true;
3564 }
3565
3566 bool CMSCollector::do_marking_st() {
3567 ResourceMark rm;
3568 HandleMark hm;
3569
3570 // Temporarily make refs discovery single threaded (non-MT)
3571 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false);
3572 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap,
3573 &_markStack, CMSYield);
3574 // the last argument to iterate indicates whether the iteration
3575 // should be incremental with periodic yields.
3576 _markBitMap.iterate(&markFromRootsClosure);
3577 // If _restart_addr is non-NULL, a marking stack overflow
3578 // occurred; we need to do a fresh iteration from the
3579 // indicated restart address.
3580 while (_restart_addr != NULL) {
3581 if (_foregroundGCIsActive) {
3582 // We may be running into repeated stack overflows, having
3583 // reached the limit of the stack size, while making very
3584 // slow forward progress. It may be best to bail out and
3585 // let the foreground collector do its job.
3586 // Clear _restart_addr, so that foreground GC
3587 // works from scratch. This avoids the headache of
3588 // a "rescan" which would otherwise be needed because
3589 // of the dirty mod union table & card table.
3590 _restart_addr = NULL;
3591 return false; // indicating failure to complete marking
3592 }
3593 // Deal with stack overflow:
3594 // we restart marking from _restart_addr
3595 HeapWord* ra = _restart_addr;
3596 markFromRootsClosure.reset(ra);
3597 _restart_addr = NULL;
3598 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end());
3599 }
3600 return true;
3601 }
3602
3603 void CMSCollector::preclean() {
3604 check_correct_thread_executing();
3605 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread");
3606 verify_work_stacks_empty();
3607 verify_overflow_empty();
3608 _abort_preclean = false;
3609 if (CMSPrecleaningEnabled) {
3610 if (!CMSEdenChunksRecordAlways) {
3611 _eden_chunk_index = 0;
3612 }
3613 size_t used = get_eden_used();
3614 size_t capacity = get_eden_capacity();
3615 // Don't start sampling unless we will get sufficiently
3616 // many samples.
3617 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100)
3618 * CMSScheduleRemarkEdenPenetration)) {
3619 _start_sampling = true;
3620 } else {
3621 _start_sampling = false;
3622 }
3623 GCTraceCPUTime tcpu;
3624 CMSPhaseAccounting pa(this, "Concurrent Preclean");
3625 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1);
3626 }
3627 CMSTokenSync x(true); // is cms thread
3628 if (CMSPrecleaningEnabled) {
3629 sample_eden();
3630 _collectorState = AbortablePreclean;
3631 } else {
3632 _collectorState = FinalMarking;
3633 }
3634 verify_work_stacks_empty();
3635 verify_overflow_empty();
3636 }
3637
3638 // Try and schedule the remark such that young gen
3639 // occupancy is CMSScheduleRemarkEdenPenetration %.
3640 void CMSCollector::abortable_preclean() {
3641 check_correct_thread_executing();
3642 assert(CMSPrecleaningEnabled, "Inconsistent control state");
3643 assert(_collectorState == AbortablePreclean, "Inconsistent control state");
3644
3645 // If Eden's current occupancy is below this threshold,
3646 // immediately schedule the remark; else preclean
3647 // past the next scavenge in an effort to
3648 // schedule the pause as described above. By choosing
3649 // CMSScheduleRemarkEdenSizeThreshold >= max eden size
3650 // we will never do an actual abortable preclean cycle.
3651 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) {
3652 GCTraceCPUTime tcpu;
3653 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean");
3654 // We need more smarts in the abortable preclean
3655 // loop below to deal with cases where allocation
3656 // in young gen is very very slow, and our precleaning
3657 // is running a losing race against a horde of
3658 // mutators intent on flooding us with CMS updates
3659 // (dirty cards).
3660 // One, admittedly dumb, strategy is to give up
3661 // after a certain number of abortable precleaning loops
3662 // or after a certain maximum time. We want to make
3663 // this smarter in the next iteration.
3664 // XXX FIX ME!!! YSR
3665 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0;
3666 while (!(should_abort_preclean() ||
3667 ConcurrentMarkSweepThread::cmst()->should_terminate())) {
3668 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2);
3669 cumworkdone += workdone;
3670 loops++;
3671 // Voluntarily terminate abortable preclean phase if we have
3672 // been at it for too long.
3673 if ((CMSMaxAbortablePrecleanLoops != 0) &&
3674 loops >= CMSMaxAbortablePrecleanLoops) {
3675 log_debug(gc)(" CMS: abort preclean due to loops ");
3676 break;
3677 }
3678 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) {
3679 log_debug(gc)(" CMS: abort preclean due to time ");
3680 break;
3681 }
3682 // If we are doing little work each iteration, we should
3683 // take a short break.
3684 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) {
3685 // Sleep for some time, waiting for work to accumulate
3686 stopTimer();
3687 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis);
3688 startTimer();
3689 waited++;
3690 }
3691 }
3692 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ",
3693 loops, waited, cumworkdone);
3694 }
3695 CMSTokenSync x(true); // is cms thread
3696 if (_collectorState != Idling) {
3697 assert(_collectorState == AbortablePreclean,
3698 "Spontaneous state transition?");
3699 _collectorState = FinalMarking;
3700 } // Else, a foreground collection completed this CMS cycle.
3701 return;
3702 }
3703
3704 // Respond to an Eden sampling opportunity
3705 void CMSCollector::sample_eden() {
3706 // Make sure a young gc cannot sneak in between our
3707 // reading and recording of a sample.
3708 assert(Thread::current()->is_ConcurrentGC_thread(),
3709 "Only the cms thread may collect Eden samples");
3710 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
3711 "Should collect samples while holding CMS token");
3712 if (!_start_sampling) {
3713 return;
3714 }
3715 // When CMSEdenChunksRecordAlways is true, the eden chunk array
3716 // is populated by the young generation.
3717 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) {
3718 if (_eden_chunk_index < _eden_chunk_capacity) {
3719 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample
3720 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
3721 "Unexpected state of Eden");
3722 // We'd like to check that what we just sampled is an oop-start address;
3723 // however, we cannot do that here since the object may not yet have been
3724 // initialized. So we'll instead do the check when we _use_ this sample
3725 // later.
3726 if (_eden_chunk_index == 0 ||
3727 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
3728 _eden_chunk_array[_eden_chunk_index-1])
3729 >= CMSSamplingGrain)) {
3730 _eden_chunk_index++; // commit sample
3731 }
3732 }
3733 }
3734 if ((_collectorState == AbortablePreclean) && !_abort_preclean) {
3735 size_t used = get_eden_used();
3736 size_t capacity = get_eden_capacity();
3737 assert(used <= capacity, "Unexpected state of Eden");
3738 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) {
3739 _abort_preclean = true;
3740 }
3741 }
3742 }
3743
3744
3745 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) {
3746 assert(_collectorState == Precleaning ||
3747 _collectorState == AbortablePreclean, "incorrect state");
3748 ResourceMark rm;
3749 HandleMark hm;
3750
3751 // Precleaning is currently not MT but the reference processor
3752 // may be set for MT. Disable it temporarily here.
3753 ReferenceProcessor* rp = ref_processor();
3754 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
3755
3756 // Do one pass of scrubbing the discovered reference lists
3757 // to remove any reference objects with strongly-reachable
3758 // referents.
3759 if (clean_refs) {
3760 CMSPrecleanRefsYieldClosure yield_cl(this);
3761 assert(rp->span().equals(_span), "Spans should be equal");
3762 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap,
3763 &_markStack, true /* preclean */);
3764 CMSDrainMarkingStackClosure complete_trace(this,
3765 _span, &_markBitMap, &_markStack,
3766 &keep_alive, true /* preclean */);
3767
3768 // We don't want this step to interfere with a young
3769 // collection because we don't want to take CPU
3770 // or memory bandwidth away from the young GC threads
3771 // (which may be as many as there are CPUs).
3772 // Note that we don't need to protect ourselves from
3773 // interference with mutators because they can't
3774 // manipulate the discovered reference lists nor affect
3775 // the computed reachability of the referents, the
3776 // only properties manipulated by the precleaning
3777 // of these reference lists.
3778 stopTimer();
3779 CMSTokenSyncWithLocks x(true /* is cms thread */,
3780 bitMapLock());
3781 startTimer();
3782 sample_eden();
3783
3784 // The following will yield to allow foreground
3785 // collection to proceed promptly. XXX YSR:
3786 // The code in this method may need further
3787 // tweaking for better performance and some restructuring
3788 // for cleaner interfaces.
3789 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases
3790 rp->preclean_discovered_references(
3791 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl,
3792 gc_timer);
3793 }
3794
3795 if (clean_survivor) { // preclean the active survivor space(s)
3796 PushAndMarkClosure pam_cl(this, _span, ref_processor(),
3797 &_markBitMap, &_modUnionTable,
3798 &_markStack, true /* precleaning phase */);
3799 stopTimer();
3800 CMSTokenSyncWithLocks ts(true /* is cms thread */,
3801 bitMapLock());
3802 startTimer();
3803 unsigned int before_count =
3804 CMSHeap::heap()->total_collections();
3805 SurvivorSpacePrecleanClosure
3806 sss_cl(this, _span, &_markBitMap, &_markStack,
3807 &pam_cl, before_count, CMSYield);
3808 _young_gen->from()->object_iterate_careful(&sss_cl);
3809 _young_gen->to()->object_iterate_careful(&sss_cl);
3810 }
3811 MarkRefsIntoAndScanClosure
3812 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable,
3813 &_markStack, this, CMSYield,
3814 true /* precleaning phase */);
3815 // CAUTION: The following closure has persistent state that may need to
3816 // be reset upon a decrease in the sequence of addresses it
3817 // processes.
3818 ScanMarkedObjectsAgainCarefullyClosure
3819 smoac_cl(this, _span,
3820 &_markBitMap, &_markStack, &mrias_cl, CMSYield);
3821
3822 // Preclean dirty cards in ModUnionTable and CardTable using
3823 // appropriate convergence criterion;
3824 // repeat CMSPrecleanIter times unless we find that
3825 // we are losing.
3826 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large");
3827 assert(CMSPrecleanNumerator < CMSPrecleanDenominator,
3828 "Bad convergence multiplier");
3829 assert(CMSPrecleanThreshold >= 100,
3830 "Unreasonably low CMSPrecleanThreshold");
3831
3832 size_t numIter, cumNumCards, lastNumCards, curNumCards;
3833 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0;
3834 numIter < CMSPrecleanIter;
3835 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) {
3836 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl);
3837 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards);
3838 // Either there are very few dirty cards, so re-mark
3839 // pause will be small anyway, or our pre-cleaning isn't
3840 // that much faster than the rate at which cards are being
3841 // dirtied, so we might as well stop and re-mark since
3842 // precleaning won't improve our re-mark time by much.
3843 if (curNumCards <= CMSPrecleanThreshold ||
3844 (numIter > 0 &&
3845 (curNumCards * CMSPrecleanDenominator >
3846 lastNumCards * CMSPrecleanNumerator))) {
3847 numIter++;
3848 cumNumCards += curNumCards;
3849 break;
3850 }
3851 }
3852
3853 preclean_cld(&mrias_cl, _cmsGen->freelistLock());
3854
3855 curNumCards = preclean_card_table(_cmsGen, &smoac_cl);
3856 cumNumCards += curNumCards;
3857 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)",
3858 curNumCards, cumNumCards, numIter);
3859 return cumNumCards; // as a measure of useful work done
3860 }
3861
3862 // PRECLEANING NOTES:
3863 // Precleaning involves:
3864 // . reading the bits of the modUnionTable and clearing the set bits.
3865 // . For the cards corresponding to the set bits, we scan the
3866 // objects on those cards. This means we need the free_list_lock
3867 // so that we can safely iterate over the CMS space when scanning
3868 // for oops.
3869 // . When we scan the objects, we'll be both reading and setting
3870 // marks in the marking bit map, so we'll need the marking bit map.
3871 // . For protecting _collector_state transitions, we take the CGC_lock.
3872 // Note that any races in the reading of of card table entries by the
3873 // CMS thread on the one hand and the clearing of those entries by the
3874 // VM thread or the setting of those entries by the mutator threads on the
3875 // other are quite benign. However, for efficiency it makes sense to keep
3876 // the VM thread from racing with the CMS thread while the latter is
3877 // dirty card info to the modUnionTable. We therefore also use the
3878 // CGC_lock to protect the reading of the card table and the mod union
3879 // table by the CM thread.
3880 // . We run concurrently with mutator updates, so scanning
3881 // needs to be done carefully -- we should not try to scan
3882 // potentially uninitialized objects.
3883 //
3884 // Locking strategy: While holding the CGC_lock, we scan over and
3885 // reset a maximal dirty range of the mod union / card tables, then lock
3886 // the free_list_lock and bitmap lock to do a full marking, then
3887 // release these locks; and repeat the cycle. This allows for a
3888 // certain amount of fairness in the sharing of these locks between
3889 // the CMS collector on the one hand, and the VM thread and the
3890 // mutators on the other.
3891
3892 // NOTE: preclean_mod_union_table() and preclean_card_table()
3893 // further below are largely identical; if you need to modify
3894 // one of these methods, please check the other method too.
3895
3896 size_t CMSCollector::preclean_mod_union_table(
3897 ConcurrentMarkSweepGeneration* old_gen,
3898 ScanMarkedObjectsAgainCarefullyClosure* cl) {
3899 verify_work_stacks_empty();
3900 verify_overflow_empty();
3901
3902 // strategy: starting with the first card, accumulate contiguous
3903 // ranges of dirty cards; clear these cards, then scan the region
3904 // covered by these cards.
3905
3906 // Since all of the MUT is committed ahead, we can just use
3907 // that, in case the generations expand while we are precleaning.
3908 // It might also be fine to just use the committed part of the
3909 // generation, but we might potentially miss cards when the
3910 // generation is rapidly expanding while we are in the midst
3911 // of precleaning.
3912 HeapWord* startAddr = old_gen->reserved().start();
3913 HeapWord* endAddr = old_gen->reserved().end();
3914
3915 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
3916
3917 size_t numDirtyCards, cumNumDirtyCards;
3918 HeapWord *nextAddr, *lastAddr;
3919 for (cumNumDirtyCards = numDirtyCards = 0,
3920 nextAddr = lastAddr = startAddr;
3921 nextAddr < endAddr;
3922 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
3923
3924 ResourceMark rm;
3925 HandleMark hm;
3926
3927 MemRegion dirtyRegion;
3928 {
3929 stopTimer();
3930 // Potential yield point
3931 CMSTokenSync ts(true);
3932 startTimer();
3933 sample_eden();
3934 // Get dirty region starting at nextOffset (inclusive),
3935 // simultaneously clearing it.
3936 dirtyRegion =
3937 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr);
3938 assert(dirtyRegion.start() >= nextAddr,
3939 "returned region inconsistent?");
3940 }
3941 // Remember where the next search should begin.
3942 // The returned region (if non-empty) is a right open interval,
3943 // so lastOffset is obtained from the right end of that
3944 // interval.
3945 lastAddr = dirtyRegion.end();
3946 // Should do something more transparent and less hacky XXX
3947 numDirtyCards =
3948 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size());
3949
3950 // We'll scan the cards in the dirty region (with periodic
3951 // yields for foreground GC as needed).
3952 if (!dirtyRegion.is_empty()) {
3953 assert(numDirtyCards > 0, "consistency check");
3954 HeapWord* stop_point = NULL;
3955 stopTimer();
3956 // Potential yield point
3957 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(),
3958 bitMapLock());
3959 startTimer();
3960 {
3961 verify_work_stacks_empty();
3962 verify_overflow_empty();
3963 sample_eden();
3964 stop_point =
3965 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
3966 }
3967 if (stop_point != NULL) {
3968 // The careful iteration stopped early either because it found an
3969 // uninitialized object, or because we were in the midst of an
3970 // "abortable preclean", which should now be aborted. Redirty
3971 // the bits corresponding to the partially-scanned or unscanned
3972 // cards. We'll either restart at the next block boundary or
3973 // abort the preclean.
3974 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
3975 "Should only be AbortablePreclean.");
3976 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end()));
3977 if (should_abort_preclean()) {
3978 break; // out of preclean loop
3979 } else {
3980 // Compute the next address at which preclean should pick up;
3981 // might need bitMapLock in order to read P-bits.
3982 lastAddr = next_card_start_after_block(stop_point);
3983 }
3984 }
3985 } else {
3986 assert(lastAddr == endAddr, "consistency check");
3987 assert(numDirtyCards == 0, "consistency check");
3988 break;
3989 }
3990 }
3991 verify_work_stacks_empty();
3992 verify_overflow_empty();
3993 return cumNumDirtyCards;
3994 }
3995
3996 // NOTE: preclean_mod_union_table() above and preclean_card_table()
3997 // below are largely identical; if you need to modify
3998 // one of these methods, please check the other method too.
3999
4000 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen,
4001 ScanMarkedObjectsAgainCarefullyClosure* cl) {
4002 // strategy: it's similar to precleamModUnionTable above, in that
4003 // we accumulate contiguous ranges of dirty cards, mark these cards
4004 // precleaned, then scan the region covered by these cards.
4005 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high());
4006 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low());
4007
4008 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding
4009
4010 size_t numDirtyCards, cumNumDirtyCards;
4011 HeapWord *lastAddr, *nextAddr;
4012
4013 for (cumNumDirtyCards = numDirtyCards = 0,
4014 nextAddr = lastAddr = startAddr;
4015 nextAddr < endAddr;
4016 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) {
4017
4018 ResourceMark rm;
4019 HandleMark hm;
4020
4021 MemRegion dirtyRegion;
4022 {
4023 // See comments in "Precleaning notes" above on why we
4024 // do this locking. XXX Could the locking overheads be
4025 // too high when dirty cards are sparse? [I don't think so.]
4026 stopTimer();
4027 CMSTokenSync x(true); // is cms thread
4028 startTimer();
4029 sample_eden();
4030 // Get and clear dirty region from card table
4031 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr),
4032 true,
4033 CardTable::precleaned_card_val());
4034
4035 assert(dirtyRegion.start() >= nextAddr,
4036 "returned region inconsistent?");
4037 }
4038 lastAddr = dirtyRegion.end();
4039 numDirtyCards =
4040 dirtyRegion.word_size()/CardTable::card_size_in_words;
4041
4042 if (!dirtyRegion.is_empty()) {
4043 stopTimer();
4044 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock());
4045 startTimer();
4046 sample_eden();
4047 verify_work_stacks_empty();
4048 verify_overflow_empty();
4049 HeapWord* stop_point =
4050 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl);
4051 if (stop_point != NULL) {
4052 assert((_collectorState == AbortablePreclean && should_abort_preclean()),
4053 "Should only be AbortablePreclean.");
4054 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end()));
4055 if (should_abort_preclean()) {
4056 break; // out of preclean loop
4057 } else {
4058 // Compute the next address at which preclean should pick up.
4059 lastAddr = next_card_start_after_block(stop_point);
4060 }
4061 }
4062 } else {
4063 break;
4064 }
4065 }
4066 verify_work_stacks_empty();
4067 verify_overflow_empty();
4068 return cumNumDirtyCards;
4069 }
4070
4071 class PrecleanCLDClosure : public CLDClosure {
4072 MetadataAwareOopsInGenClosure* _cm_closure;
4073 public:
4074 PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {}
4075 void do_cld(ClassLoaderData* cld) {
4076 if (cld->has_accumulated_modified_oops()) {
4077 cld->clear_accumulated_modified_oops();
4078
4079 _cm_closure->do_cld(cld);
4080 }
4081 }
4082 };
4083
4084 // The freelist lock is needed to prevent asserts, is it really needed?
4085 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) {
4086
4087 cl->set_freelistLock(freelistLock);
4088
4089 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock());
4090
4091 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean?
4092 // SSS: We should probably check if precleaning should be aborted, at suitable intervals?
4093 PrecleanCLDClosure preclean_closure(cl);
4094 ClassLoaderDataGraph::cld_do(&preclean_closure);
4095
4096 verify_work_stacks_empty();
4097 verify_overflow_empty();
4098 }
4099
4100 void CMSCollector::checkpointRootsFinal() {
4101 assert(_collectorState == FinalMarking, "incorrect state transition?");
4102 check_correct_thread_executing();
4103 // world is stopped at this checkpoint
4104 assert(SafepointSynchronize::is_at_safepoint(),
4105 "world should be stopped");
4106 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
4107
4108 verify_work_stacks_empty();
4109 verify_overflow_empty();
4110
4111 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)",
4112 _young_gen->used() / K, _young_gen->capacity() / K);
4113 {
4114 if (CMSScavengeBeforeRemark) {
4115 CMSHeap* heap = CMSHeap::heap();
4116 // Temporarily set flag to false, GCH->do_collection will
4117 // expect it to be false and set to true
4118 FlagSetting fl(heap->_is_gc_active, false);
4119
4120 heap->do_collection(true, // full (i.e. force, see below)
4121 false, // !clear_all_soft_refs
4122 0, // size
4123 false, // is_tlab
4124 GenCollectedHeap::YoungGen // type
4125 );
4126 }
4127 FreelistLocker x(this);
4128 MutexLockerEx y(bitMapLock(),
4129 Mutex::_no_safepoint_check_flag);
4130 checkpointRootsFinalWork();
4131 }
4132 verify_work_stacks_empty();
4133 verify_overflow_empty();
4134 }
4135
4136 void CMSCollector::checkpointRootsFinalWork() {
4137 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm);
4138
4139 assert(haveFreelistLocks(), "must have free list locks");
4140 assert_lock_strong(bitMapLock());
4141
4142 ResourceMark rm;
4143 HandleMark hm;
4144
4145 CMSHeap* heap = CMSHeap::heap();
4146
4147 if (should_unload_classes()) {
4148 CodeCache::gc_prologue();
4149 }
4150 assert(haveFreelistLocks(), "must have free list locks");
4151 assert_lock_strong(bitMapLock());
4152
4153 // We might assume that we need not fill TLAB's when
4154 // CMSScavengeBeforeRemark is set, because we may have just done
4155 // a scavenge which would have filled all TLAB's -- and besides
4156 // Eden would be empty. This however may not always be the case --
4157 // for instance although we asked for a scavenge, it may not have
4158 // happened because of a JNI critical section. We probably need
4159 // a policy for deciding whether we can in that case wait until
4160 // the critical section releases and then do the remark following
4161 // the scavenge, and skip it here. In the absence of that policy,
4162 // or of an indication of whether the scavenge did indeed occur,
4163 // we cannot rely on TLAB's having been filled and must do
4164 // so here just in case a scavenge did not happen.
4165 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them
4166 // Update the saved marks which may affect the root scans.
4167 heap->save_marks();
4168
4169 print_eden_and_survivor_chunk_arrays();
4170
4171 {
4172 #if COMPILER2_OR_JVMCI
4173 DerivedPointerTableDeactivate dpt_deact;
4174 #endif
4175
4176 // Note on the role of the mod union table:
4177 // Since the marker in "markFromRoots" marks concurrently with
4178 // mutators, it is possible for some reachable objects not to have been
4179 // scanned. For instance, an only reference to an object A was
4180 // placed in object B after the marker scanned B. Unless B is rescanned,
4181 // A would be collected. Such updates to references in marked objects
4182 // are detected via the mod union table which is the set of all cards
4183 // dirtied since the first checkpoint in this GC cycle and prior to
4184 // the most recent young generation GC, minus those cleaned up by the
4185 // concurrent precleaning.
4186 if (CMSParallelRemarkEnabled) {
4187 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm);
4188 do_remark_parallel();
4189 } else {
4190 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm);
4191 do_remark_non_parallel();
4192 }
4193 }
4194 verify_work_stacks_empty();
4195 verify_overflow_empty();
4196
4197 {
4198 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm);
4199 refProcessingWork();
4200 }
4201 verify_work_stacks_empty();
4202 verify_overflow_empty();
4203
4204 if (should_unload_classes()) {
4205 CodeCache::gc_epilogue();
4206 }
4207 JvmtiExport::gc_epilogue();
4208
4209 // If we encountered any (marking stack / work queue) overflow
4210 // events during the current CMS cycle, take appropriate
4211 // remedial measures, where possible, so as to try and avoid
4212 // recurrence of that condition.
4213 assert(_markStack.isEmpty(), "No grey objects");
4214 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw +
4215 _ser_kac_ovflw + _ser_kac_preclean_ovflw;
4216 if (ser_ovflw > 0) {
4217 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")",
4218 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw);
4219 _markStack.expand();
4220 _ser_pmc_remark_ovflw = 0;
4221 _ser_pmc_preclean_ovflw = 0;
4222 _ser_kac_preclean_ovflw = 0;
4223 _ser_kac_ovflw = 0;
4224 }
4225 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) {
4226 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")",
4227 _par_pmc_remark_ovflw, _par_kac_ovflw);
4228 _par_pmc_remark_ovflw = 0;
4229 _par_kac_ovflw = 0;
4230 }
4231 if (_markStack._hit_limit > 0) {
4232 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")",
4233 _markStack._hit_limit);
4234 }
4235 if (_markStack._failed_double > 0) {
4236 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT,
4237 _markStack._failed_double, _markStack.capacity());
4238 }
4239 _markStack._hit_limit = 0;
4240 _markStack._failed_double = 0;
4241
4242 if ((VerifyAfterGC || VerifyDuringGC) &&
4243 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4244 verify_after_remark();
4245 }
4246
4247 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure);
4248
4249 // Change under the freelistLocks.
4250 _collectorState = Sweeping;
4251 // Call isAllClear() under bitMapLock
4252 assert(_modUnionTable.isAllClear(),
4253 "Should be clear by end of the final marking");
4254 assert(_ct->cld_rem_set()->mod_union_is_clear(),
4255 "Should be clear by end of the final marking");
4256 }
4257
4258 void CMSParInitialMarkTask::work(uint worker_id) {
4259 elapsedTimer _timer;
4260 ResourceMark rm;
4261 HandleMark hm;
4262
4263 // ---------- scan from roots --------------
4264 _timer.start();
4265 CMSHeap* heap = CMSHeap::heap();
4266 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap));
4267
4268 // ---------- young gen roots --------------
4269 {
4270 work_on_young_gen_roots(&par_mri_cl);
4271 _timer.stop();
4272 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4273 }
4274
4275 // ---------- remaining roots --------------
4276 _timer.reset();
4277 _timer.start();
4278
4279 CLDToOopClosure cld_closure(&par_mri_cl, true);
4280
4281 heap->cms_process_roots(_strong_roots_scope,
4282 false, // yg was scanned above
4283 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4284 _collector->should_unload_classes(),
4285 &par_mri_cl,
4286 &cld_closure);
4287 assert(_collector->should_unload_classes()
4288 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4289 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4290 _timer.stop();
4291 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4292 }
4293
4294 // Parallel remark task
4295 class CMSParRemarkTask: public CMSParMarkTask {
4296 CompactibleFreeListSpace* _cms_space;
4297
4298 // The per-thread work queues, available here for stealing.
4299 OopTaskQueueSet* _task_queues;
4300 ParallelTaskTerminator _term;
4301 StrongRootsScope* _strong_roots_scope;
4302
4303 public:
4304 // A value of 0 passed to n_workers will cause the number of
4305 // workers to be taken from the active workers in the work gang.
4306 CMSParRemarkTask(CMSCollector* collector,
4307 CompactibleFreeListSpace* cms_space,
4308 uint n_workers, WorkGang* workers,
4309 OopTaskQueueSet* task_queues,
4310 StrongRootsScope* strong_roots_scope):
4311 CMSParMarkTask("Rescan roots and grey objects in parallel",
4312 collector, n_workers),
4313 _cms_space(cms_space),
4314 _task_queues(task_queues),
4315 _term(n_workers, task_queues),
4316 _strong_roots_scope(strong_roots_scope) { }
4317
4318 OopTaskQueueSet* task_queues() { return _task_queues; }
4319
4320 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
4321
4322 ParallelTaskTerminator* terminator() { return &_term; }
4323 uint n_workers() { return _n_workers; }
4324
4325 void work(uint worker_id);
4326
4327 private:
4328 // ... of dirty cards in old space
4329 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i,
4330 ParMarkRefsIntoAndScanClosure* cl);
4331
4332 // ... work stealing for the above
4333 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed);
4334 };
4335
4336 class RemarkCLDClosure : public CLDClosure {
4337 CLDToOopClosure _cm_closure;
4338 public:
4339 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {}
4340 void do_cld(ClassLoaderData* cld) {
4341 // Check if we have modified any oops in the CLD during the concurrent marking.
4342 if (cld->has_accumulated_modified_oops()) {
4343 cld->clear_accumulated_modified_oops();
4344
4345 // We could have transfered the current modified marks to the accumulated marks,
4346 // like we do with the Card Table to Mod Union Table. But it's not really necessary.
4347 } else if (cld->has_modified_oops()) {
4348 // Don't clear anything, this info is needed by the next young collection.
4349 } else {
4350 // No modified oops in the ClassLoaderData.
4351 return;
4352 }
4353
4354 // The klass has modified fields, need to scan the klass.
4355 _cm_closure.do_cld(cld);
4356 }
4357 };
4358
4359 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) {
4360 ParNewGeneration* young_gen = _collector->_young_gen;
4361 ContiguousSpace* eden_space = young_gen->eden();
4362 ContiguousSpace* from_space = young_gen->from();
4363 ContiguousSpace* to_space = young_gen->to();
4364
4365 HeapWord** eca = _collector->_eden_chunk_array;
4366 size_t ect = _collector->_eden_chunk_index;
4367 HeapWord** sca = _collector->_survivor_chunk_array;
4368 size_t sct = _collector->_survivor_chunk_index;
4369
4370 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds");
4371 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds");
4372
4373 do_young_space_rescan(cl, to_space, NULL, 0);
4374 do_young_space_rescan(cl, from_space, sca, sct);
4375 do_young_space_rescan(cl, eden_space, eca, ect);
4376 }
4377
4378 // work_queue(i) is passed to the closure
4379 // ParMarkRefsIntoAndScanClosure. The "i" parameter
4380 // also is passed to do_dirty_card_rescan_tasks() and to
4381 // do_work_steal() to select the i-th task_queue.
4382
4383 void CMSParRemarkTask::work(uint worker_id) {
4384 elapsedTimer _timer;
4385 ResourceMark rm;
4386 HandleMark hm;
4387
4388 // ---------- rescan from roots --------------
4389 _timer.start();
4390 CMSHeap* heap = CMSHeap::heap();
4391 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector,
4392 _collector->_span, _collector->ref_processor(),
4393 &(_collector->_markBitMap),
4394 work_queue(worker_id));
4395
4396 // Rescan young gen roots first since these are likely
4397 // coarsely partitioned and may, on that account, constitute
4398 // the critical path; thus, it's best to start off that
4399 // work first.
4400 // ---------- young gen roots --------------
4401 {
4402 work_on_young_gen_roots(&par_mrias_cl);
4403 _timer.stop();
4404 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4405 }
4406
4407 // ---------- remaining roots --------------
4408 _timer.reset();
4409 _timer.start();
4410 heap->cms_process_roots(_strong_roots_scope,
4411 false, // yg was scanned above
4412 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()),
4413 _collector->should_unload_classes(),
4414 &par_mrias_cl,
4415 NULL); // The dirty klasses will be handled below
4416
4417 assert(_collector->should_unload_classes()
4418 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4419 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4420 _timer.stop();
4421 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4422
4423 // ---------- unhandled CLD scanning ----------
4424 if (worker_id == 0) { // Single threaded at the moment.
4425 _timer.reset();
4426 _timer.start();
4427
4428 // Scan all new class loader data objects and new dependencies that were
4429 // introduced during concurrent marking.
4430 ResourceMark rm;
4431 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4432 for (int i = 0; i < array->length(); i++) {
4433 par_mrias_cl.do_cld_nv(array->at(i));
4434 }
4435
4436 // We don't need to keep track of new CLDs anymore.
4437 ClassLoaderDataGraph::remember_new_clds(false);
4438
4439 _timer.stop();
4440 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4441 }
4442
4443 // We might have added oops to ClassLoaderData::_handles during the
4444 // concurrent marking phase. These oops do not always point to newly allocated objects
4445 // that are guaranteed to be kept alive. Hence,
4446 // we do have to revisit the _handles block during the remark phase.
4447
4448 // ---------- dirty CLD scanning ----------
4449 if (worker_id == 0) { // Single threaded at the moment.
4450 _timer.reset();
4451 _timer.start();
4452
4453 // Scan all classes that was dirtied during the concurrent marking phase.
4454 RemarkCLDClosure remark_closure(&par_mrias_cl);
4455 ClassLoaderDataGraph::cld_do(&remark_closure);
4456
4457 _timer.stop();
4458 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4459 }
4460
4461
4462 // ---------- rescan dirty cards ------------
4463 _timer.reset();
4464 _timer.start();
4465
4466 // Do the rescan tasks for each of the two spaces
4467 // (cms_space) in turn.
4468 // "worker_id" is passed to select the task_queue for "worker_id"
4469 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl);
4470 _timer.stop();
4471 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4472
4473 // ---------- steal work from other threads ...
4474 // ---------- ... and drain overflow list.
4475 _timer.reset();
4476 _timer.start();
4477 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
4478 _timer.stop();
4479 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds());
4480 }
4481
4482 void
4483 CMSParMarkTask::do_young_space_rescan(
4484 OopsInGenClosure* cl, ContiguousSpace* space,
4485 HeapWord** chunk_array, size_t chunk_top) {
4486 // Until all tasks completed:
4487 // . claim an unclaimed task
4488 // . compute region boundaries corresponding to task claimed
4489 // using chunk_array
4490 // . par_oop_iterate(cl) over that region
4491
4492 ResourceMark rm;
4493 HandleMark hm;
4494
4495 SequentialSubTasksDone* pst = space->par_seq_tasks();
4496
4497 uint nth_task = 0;
4498 uint n_tasks = pst->n_tasks();
4499
4500 if (n_tasks > 0) {
4501 assert(pst->valid(), "Uninitialized use?");
4502 HeapWord *start, *end;
4503 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4504 // We claimed task # nth_task; compute its boundaries.
4505 if (chunk_top == 0) { // no samples were taken
4506 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task");
4507 start = space->bottom();
4508 end = space->top();
4509 } else if (nth_task == 0) {
4510 start = space->bottom();
4511 end = chunk_array[nth_task];
4512 } else if (nth_task < (uint)chunk_top) {
4513 assert(nth_task >= 1, "Control point invariant");
4514 start = chunk_array[nth_task - 1];
4515 end = chunk_array[nth_task];
4516 } else {
4517 assert(nth_task == (uint)chunk_top, "Control point invariant");
4518 start = chunk_array[chunk_top - 1];
4519 end = space->top();
4520 }
4521 MemRegion mr(start, end);
4522 // Verify that mr is in space
4523 assert(mr.is_empty() || space->used_region().contains(mr),
4524 "Should be in space");
4525 // Verify that "start" is an object boundary
4526 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())),
4527 "Should be an oop");
4528 space->par_oop_iterate(mr, cl);
4529 }
4530 pst->all_tasks_completed();
4531 }
4532 }
4533
4534 void
4535 CMSParRemarkTask::do_dirty_card_rescan_tasks(
4536 CompactibleFreeListSpace* sp, int i,
4537 ParMarkRefsIntoAndScanClosure* cl) {
4538 // Until all tasks completed:
4539 // . claim an unclaimed task
4540 // . compute region boundaries corresponding to task claimed
4541 // . transfer dirty bits ct->mut for that region
4542 // . apply rescanclosure to dirty mut bits for that region
4543
4544 ResourceMark rm;
4545 HandleMark hm;
4546
4547 OopTaskQueue* work_q = work_queue(i);
4548 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable));
4549 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION!
4550 // CAUTION: This closure has state that persists across calls to
4551 // the work method dirty_range_iterate_clear() in that it has
4552 // embedded in it a (subtype of) UpwardsObjectClosure. The
4553 // use of that state in the embedded UpwardsObjectClosure instance
4554 // assumes that the cards are always iterated (even if in parallel
4555 // by several threads) in monotonically increasing order per each
4556 // thread. This is true of the implementation below which picks
4557 // card ranges (chunks) in monotonically increasing order globally
4558 // and, a-fortiori, in monotonically increasing order per thread
4559 // (the latter order being a subsequence of the former).
4560 // If the work code below is ever reorganized into a more chaotic
4561 // work-partitioning form than the current "sequential tasks"
4562 // paradigm, the use of that persistent state will have to be
4563 // revisited and modified appropriately. See also related
4564 // bug 4756801 work on which should examine this code to make
4565 // sure that the changes there do not run counter to the
4566 // assumptions made here and necessary for correctness and
4567 // efficiency. Note also that this code might yield inefficient
4568 // behavior in the case of very large objects that span one or
4569 // more work chunks. Such objects would potentially be scanned
4570 // several times redundantly. Work on 4756801 should try and
4571 // address that performance anomaly if at all possible. XXX
4572 MemRegion full_span = _collector->_span;
4573 CMSBitMap* bm = &(_collector->_markBitMap); // shared
4574 MarkFromDirtyCardsClosure
4575 greyRescanClosure(_collector, full_span, // entire span of interest
4576 sp, bm, work_q, cl);
4577
4578 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks();
4579 assert(pst->valid(), "Uninitialized use?");
4580 uint nth_task = 0;
4581 const int alignment = CardTable::card_size * BitsPerWord;
4582 MemRegion span = sp->used_region();
4583 HeapWord* start_addr = span.start();
4584 HeapWord* end_addr = align_up(span.end(), alignment);
4585 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units
4586 assert(is_aligned(start_addr, alignment), "Check alignment");
4587 assert(is_aligned(chunk_size, alignment), "Check alignment");
4588
4589 while (!pst->is_task_claimed(/* reference */ nth_task)) {
4590 // Having claimed the nth_task, compute corresponding mem-region,
4591 // which is a-fortiori aligned correctly (i.e. at a MUT boundary).
4592 // The alignment restriction ensures that we do not need any
4593 // synchronization with other gang-workers while setting or
4594 // clearing bits in thus chunk of the MUT.
4595 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size,
4596 start_addr + (nth_task+1)*chunk_size);
4597 // The last chunk's end might be way beyond end of the
4598 // used region. In that case pull back appropriately.
4599 if (this_span.end() > end_addr) {
4600 this_span.set_end(end_addr);
4601 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)");
4602 }
4603 // Iterate over the dirty cards covering this chunk, marking them
4604 // precleaned, and setting the corresponding bits in the mod union
4605 // table. Since we have been careful to partition at Card and MUT-word
4606 // boundaries no synchronization is needed between parallel threads.
4607 _collector->_ct->dirty_card_iterate(this_span,
4608 &modUnionClosure);
4609
4610 // Having transferred these marks into the modUnionTable,
4611 // rescan the marked objects on the dirty cards in the modUnionTable.
4612 // Even if this is at a synchronous collection, the initial marking
4613 // may have been done during an asynchronous collection so there
4614 // may be dirty bits in the mod-union table.
4615 _collector->_modUnionTable.dirty_range_iterate_clear(
4616 this_span, &greyRescanClosure);
4617 _collector->_modUnionTable.verifyNoOneBitsInRange(
4618 this_span.start(),
4619 this_span.end());
4620 }
4621 pst->all_tasks_completed(); // declare that i am done
4622 }
4623
4624 // . see if we can share work_queues with ParNew? XXX
4625 void
4626 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl,
4627 int* seed) {
4628 OopTaskQueue* work_q = work_queue(i);
4629 NOT_PRODUCT(int num_steals = 0;)
4630 oop obj_to_scan;
4631 CMSBitMap* bm = &(_collector->_markBitMap);
4632
4633 while (true) {
4634 // Completely finish any left over work from (an) earlier round(s)
4635 cl->trim_queue(0);
4636 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
4637 (size_t)ParGCDesiredObjsFromOverflowList);
4638 // Now check if there's any work in the overflow list
4639 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
4640 // only affects the number of attempts made to get work from the
4641 // overflow list and does not affect the number of workers. Just
4642 // pass ParallelGCThreads so this behavior is unchanged.
4643 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
4644 work_q,
4645 ParallelGCThreads)) {
4646 // found something in global overflow list;
4647 // not yet ready to go stealing work from others.
4648 // We'd like to assert(work_q->size() != 0, ...)
4649 // because we just took work from the overflow list,
4650 // but of course we can't since all of that could have
4651 // been already stolen from us.
4652 // "He giveth and He taketh away."
4653 continue;
4654 }
4655 // Verify that we have no work before we resort to stealing
4656 assert(work_q->size() == 0, "Have work, shouldn't steal");
4657 // Try to steal from other queues that have work
4658 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
4659 NOT_PRODUCT(num_steals++;)
4660 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
4661 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
4662 // Do scanning work
4663 obj_to_scan->oop_iterate(cl);
4664 // Loop around, finish this work, and try to steal some more
4665 } else if (terminator()->offer_termination()) {
4666 break; // nirvana from the infinite cycle
4667 }
4668 }
4669 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
4670 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(),
4671 "Else our work is not yet done");
4672 }
4673
4674 // Record object boundaries in _eden_chunk_array by sampling the eden
4675 // top in the slow-path eden object allocation code path and record
4676 // the boundaries, if CMSEdenChunksRecordAlways is true. If
4677 // CMSEdenChunksRecordAlways is false, we use the other asynchronous
4678 // sampling in sample_eden() that activates during the part of the
4679 // preclean phase.
4680 void CMSCollector::sample_eden_chunk() {
4681 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) {
4682 if (_eden_chunk_lock->try_lock()) {
4683 // Record a sample. This is the critical section. The contents
4684 // of the _eden_chunk_array have to be non-decreasing in the
4685 // address order.
4686 _eden_chunk_array[_eden_chunk_index] = *_top_addr;
4687 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr,
4688 "Unexpected state of Eden");
4689 if (_eden_chunk_index == 0 ||
4690 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) &&
4691 (pointer_delta(_eden_chunk_array[_eden_chunk_index],
4692 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) {
4693 _eden_chunk_index++; // commit sample
4694 }
4695 _eden_chunk_lock->unlock();
4696 }
4697 }
4698 }
4699
4700 // Return a thread-local PLAB recording array, as appropriate.
4701 void* CMSCollector::get_data_recorder(int thr_num) {
4702 if (_survivor_plab_array != NULL &&
4703 (CMSPLABRecordAlways ||
4704 (_collectorState > Marking && _collectorState < FinalMarking))) {
4705 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds");
4706 ChunkArray* ca = &_survivor_plab_array[thr_num];
4707 ca->reset(); // clear it so that fresh data is recorded
4708 return (void*) ca;
4709 } else {
4710 return NULL;
4711 }
4712 }
4713
4714 // Reset all the thread-local PLAB recording arrays
4715 void CMSCollector::reset_survivor_plab_arrays() {
4716 for (uint i = 0; i < ParallelGCThreads; i++) {
4717 _survivor_plab_array[i].reset();
4718 }
4719 }
4720
4721 // Merge the per-thread plab arrays into the global survivor chunk
4722 // array which will provide the partitioning of the survivor space
4723 // for CMS initial scan and rescan.
4724 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
4725 int no_of_gc_threads) {
4726 assert(_survivor_plab_array != NULL, "Error");
4727 assert(_survivor_chunk_array != NULL, "Error");
4728 assert(_collectorState == FinalMarking ||
4729 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error");
4730 for (int j = 0; j < no_of_gc_threads; j++) {
4731 _cursor[j] = 0;
4732 }
4733 HeapWord* top = surv->top();
4734 size_t i;
4735 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries
4736 HeapWord* min_val = top; // Higher than any PLAB address
4737 uint min_tid = 0; // position of min_val this round
4738 for (int j = 0; j < no_of_gc_threads; j++) {
4739 ChunkArray* cur_sca = &_survivor_plab_array[j];
4740 if (_cursor[j] == cur_sca->end()) {
4741 continue;
4742 }
4743 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant");
4744 HeapWord* cur_val = cur_sca->nth(_cursor[j]);
4745 assert(surv->used_region().contains(cur_val), "Out of bounds value");
4746 if (cur_val < min_val) {
4747 min_tid = j;
4748 min_val = cur_val;
4749 } else {
4750 assert(cur_val < top, "All recorded addresses should be less");
4751 }
4752 }
4753 // At this point min_val and min_tid are respectively
4754 // the least address in _survivor_plab_array[j]->nth(_cursor[j])
4755 // and the thread (j) that witnesses that address.
4756 // We record this address in the _survivor_chunk_array[i]
4757 // and increment _cursor[min_tid] prior to the next round i.
4758 if (min_val == top) {
4759 break;
4760 }
4761 _survivor_chunk_array[i] = min_val;
4762 _cursor[min_tid]++;
4763 }
4764 // We are all done; record the size of the _survivor_chunk_array
4765 _survivor_chunk_index = i; // exclusive: [0, i)
4766 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i);
4767 // Verify that we used up all the recorded entries
4768 #ifdef ASSERT
4769 size_t total = 0;
4770 for (int j = 0; j < no_of_gc_threads; j++) {
4771 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
4772 total += _cursor[j];
4773 }
4774 assert(total == _survivor_chunk_index, "Ctl Pt Invariant");
4775 // Check that the merged array is in sorted order
4776 if (total > 0) {
4777 for (size_t i = 0; i < total - 1; i++) {
4778 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ",
4779 i, p2i(_survivor_chunk_array[i]));
4780 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1],
4781 "Not sorted");
4782 }
4783 }
4784 #endif // ASSERT
4785 }
4786
4787 // Set up the space's par_seq_tasks structure for work claiming
4788 // for parallel initial scan and rescan of young gen.
4789 // See ParRescanTask where this is currently used.
4790 void
4791 CMSCollector::
4792 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) {
4793 assert(n_threads > 0, "Unexpected n_threads argument");
4794
4795 // Eden space
4796 if (!_young_gen->eden()->is_empty()) {
4797 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks();
4798 assert(!pst->valid(), "Clobbering existing data?");
4799 // Each valid entry in [0, _eden_chunk_index) represents a task.
4800 size_t n_tasks = _eden_chunk_index + 1;
4801 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
4802 // Sets the condition for completion of the subtask (how many threads
4803 // need to finish in order to be done).
4804 pst->set_n_threads(n_threads);
4805 pst->set_n_tasks((int)n_tasks);
4806 }
4807
4808 // Merge the survivor plab arrays into _survivor_chunk_array
4809 if (_survivor_plab_array != NULL) {
4810 merge_survivor_plab_arrays(_young_gen->from(), n_threads);
4811 } else {
4812 assert(_survivor_chunk_index == 0, "Error");
4813 }
4814
4815 // To space
4816 {
4817 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks();
4818 assert(!pst->valid(), "Clobbering existing data?");
4819 // Sets the condition for completion of the subtask (how many threads
4820 // need to finish in order to be done).
4821 pst->set_n_threads(n_threads);
4822 pst->set_n_tasks(1);
4823 assert(pst->valid(), "Error");
4824 }
4825
4826 // From space
4827 {
4828 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks();
4829 assert(!pst->valid(), "Clobbering existing data?");
4830 size_t n_tasks = _survivor_chunk_index + 1;
4831 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
4832 // Sets the condition for completion of the subtask (how many threads
4833 // need to finish in order to be done).
4834 pst->set_n_threads(n_threads);
4835 pst->set_n_tasks((int)n_tasks);
4836 assert(pst->valid(), "Error");
4837 }
4838 }
4839
4840 // Parallel version of remark
4841 void CMSCollector::do_remark_parallel() {
4842 CMSHeap* heap = CMSHeap::heap();
4843 WorkGang* workers = heap->workers();
4844 assert(workers != NULL, "Need parallel worker threads.");
4845 // Choose to use the number of GC workers most recently set
4846 // into "active_workers".
4847 uint n_workers = workers->active_workers();
4848
4849 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace();
4850
4851 StrongRootsScope srs(n_workers);
4852
4853 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs);
4854
4855 // We won't be iterating over the cards in the card table updating
4856 // the younger_gen cards, so we shouldn't call the following else
4857 // the verification code as well as subsequent younger_refs_iterate
4858 // code would get confused. XXX
4859 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel
4860
4861 // The young gen rescan work will not be done as part of
4862 // process_roots (which currently doesn't know how to
4863 // parallelize such a scan), but rather will be broken up into
4864 // a set of parallel tasks (via the sampling that the [abortable]
4865 // preclean phase did of eden, plus the [two] tasks of
4866 // scanning the [two] survivor spaces. Further fine-grain
4867 // parallelization of the scanning of the survivor spaces
4868 // themselves, and of precleaning of the young gen itself
4869 // is deferred to the future.
4870 initialize_sequential_subtasks_for_young_gen_rescan(n_workers);
4871
4872 // The dirty card rescan work is broken up into a "sequence"
4873 // of parallel tasks (per constituent space) that are dynamically
4874 // claimed by the parallel threads.
4875 cms_space->initialize_sequential_subtasks_for_rescan(n_workers);
4876
4877 // It turns out that even when we're using 1 thread, doing the work in a
4878 // separate thread causes wide variance in run times. We can't help this
4879 // in the multi-threaded case, but we special-case n=1 here to get
4880 // repeatable measurements of the 1-thread overhead of the parallel code.
4881 if (n_workers > 1) {
4882 // Make refs discovery MT-safe, if it isn't already: it may not
4883 // necessarily be so, since it's possible that we are doing
4884 // ST marking.
4885 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true);
4886 workers->run_task(&tsk);
4887 } else {
4888 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4889 tsk.work(0);
4890 }
4891
4892 // restore, single-threaded for now, any preserved marks
4893 // as a result of work_q overflow
4894 restore_preserved_marks_if_any();
4895 }
4896
4897 // Non-parallel version of remark
4898 void CMSCollector::do_remark_non_parallel() {
4899 ResourceMark rm;
4900 HandleMark hm;
4901 CMSHeap* heap = CMSHeap::heap();
4902 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false);
4903
4904 MarkRefsIntoAndScanClosure
4905 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */,
4906 &_markStack, this,
4907 false /* should_yield */, false /* not precleaning */);
4908 MarkFromDirtyCardsClosure
4909 markFromDirtyCardsClosure(this, _span,
4910 NULL, // space is set further below
4911 &_markBitMap, &_markStack, &mrias_cl);
4912 {
4913 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm);
4914 // Iterate over the dirty cards, setting the corresponding bits in the
4915 // mod union table.
4916 {
4917 ModUnionClosure modUnionClosure(&_modUnionTable);
4918 _ct->dirty_card_iterate(_cmsGen->used_region(),
4919 &modUnionClosure);
4920 }
4921 // Having transferred these marks into the modUnionTable, we just need
4922 // to rescan the marked objects on the dirty cards in the modUnionTable.
4923 // The initial marking may have been done during an asynchronous
4924 // collection so there may be dirty bits in the mod-union table.
4925 const int alignment = CardTable::card_size * BitsPerWord;
4926 {
4927 // ... First handle dirty cards in CMS gen
4928 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace());
4929 MemRegion ur = _cmsGen->used_region();
4930 HeapWord* lb = ur.start();
4931 HeapWord* ub = align_up(ur.end(), alignment);
4932 MemRegion cms_span(lb, ub);
4933 _modUnionTable.dirty_range_iterate_clear(cms_span,
4934 &markFromDirtyCardsClosure);
4935 verify_work_stacks_empty();
4936 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards());
4937 }
4938 }
4939 if (VerifyDuringGC &&
4940 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) {
4941 HandleMark hm; // Discard invalid handles created during verification
4942 Universe::verify();
4943 }
4944 {
4945 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm);
4946
4947 verify_work_stacks_empty();
4948
4949 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel.
4950 StrongRootsScope srs(1);
4951
4952 heap->cms_process_roots(&srs,
4953 true, // young gen as roots
4954 GenCollectedHeap::ScanningOption(roots_scanning_options()),
4955 should_unload_classes(),
4956 &mrias_cl,
4957 NULL); // The dirty klasses will be handled below
4958
4959 assert(should_unload_classes()
4960 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache),
4961 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops");
4962 }
4963
4964 {
4965 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm);
4966
4967 verify_work_stacks_empty();
4968
4969 // Scan all class loader data objects that might have been introduced
4970 // during concurrent marking.
4971 ResourceMark rm;
4972 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds();
4973 for (int i = 0; i < array->length(); i++) {
4974 mrias_cl.do_cld_nv(array->at(i));
4975 }
4976
4977 // We don't need to keep track of new CLDs anymore.
4978 ClassLoaderDataGraph::remember_new_clds(false);
4979
4980 verify_work_stacks_empty();
4981 }
4982
4983 // We might have added oops to ClassLoaderData::_handles during the
4984 // concurrent marking phase. These oops do not point to newly allocated objects
4985 // that are guaranteed to be kept alive. Hence,
4986 // we do have to revisit the _handles block during the remark phase.
4987 {
4988 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm);
4989
4990 verify_work_stacks_empty();
4991
4992 RemarkCLDClosure remark_closure(&mrias_cl);
4993 ClassLoaderDataGraph::cld_do(&remark_closure);
4994
4995 verify_work_stacks_empty();
4996 }
4997
4998 verify_work_stacks_empty();
4999 // Restore evacuated mark words, if any, used for overflow list links
5000 restore_preserved_marks_if_any();
5001
5002 verify_overflow_empty();
5003 }
5004
5005 ////////////////////////////////////////////////////////
5006 // Parallel Reference Processing Task Proxy Class
5007 ////////////////////////////////////////////////////////
5008 class AbstractGangTaskWOopQueues : public AbstractGangTask {
5009 OopTaskQueueSet* _queues;
5010 ParallelTaskTerminator _terminator;
5011 public:
5012 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) :
5013 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {}
5014 ParallelTaskTerminator* terminator() { return &_terminator; }
5015 OopTaskQueueSet* queues() { return _queues; }
5016 };
5017
5018 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
5019 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5020 CMSCollector* _collector;
5021 CMSBitMap* _mark_bit_map;
5022 const MemRegion _span;
5023 ProcessTask& _task;
5024
5025 public:
5026 CMSRefProcTaskProxy(ProcessTask& task,
5027 CMSCollector* collector,
5028 const MemRegion& span,
5029 CMSBitMap* mark_bit_map,
5030 AbstractWorkGang* workers,
5031 OopTaskQueueSet* task_queues):
5032 AbstractGangTaskWOopQueues("Process referents by policy in parallel",
5033 task_queues,
5034 workers->active_workers()),
5035 _task(task),
5036 _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
5037 {
5038 assert(_collector->_span.equals(_span) && !_span.is_empty(),
5039 "Inconsistency in _span");
5040 }
5041
5042 OopTaskQueueSet* task_queues() { return queues(); }
5043
5044 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
5045
5046 void do_work_steal(int i,
5047 CMSParDrainMarkingStackClosure* drain,
5048 CMSParKeepAliveClosure* keep_alive,
5049 int* seed);
5050
5051 virtual void work(uint worker_id);
5052 };
5053
5054 void CMSRefProcTaskProxy::work(uint worker_id) {
5055 ResourceMark rm;
5056 HandleMark hm;
5057 assert(_collector->_span.equals(_span), "Inconsistency in _span");
5058 CMSParKeepAliveClosure par_keep_alive(_collector, _span,
5059 _mark_bit_map,
5060 work_queue(worker_id));
5061 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span,
5062 _mark_bit_map,
5063 work_queue(worker_id));
5064 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
5065 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
5066 if (_task.marks_oops_alive()) {
5067 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
5068 _collector->hash_seed(worker_id));
5069 }
5070 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
5071 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
5072 }
5073
5074 class CMSRefEnqueueTaskProxy: public AbstractGangTask {
5075 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5076 EnqueueTask& _task;
5077
5078 public:
5079 CMSRefEnqueueTaskProxy(EnqueueTask& task)
5080 : AbstractGangTask("Enqueue reference objects in parallel"),
5081 _task(task)
5082 { }
5083
5084 virtual void work(uint worker_id)
5085 {
5086 _task.work(worker_id);
5087 }
5088 };
5089
5090 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
5091 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue):
5092 _span(span),
5093 _bit_map(bit_map),
5094 _work_queue(work_queue),
5095 _mark_and_push(collector, span, bit_map, work_queue),
5096 _low_water_mark(MIN2((work_queue->max_elems()/4),
5097 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads)))
5098 { }
5099
5100 // . see if we can share work_queues with ParNew? XXX
5101 void CMSRefProcTaskProxy::do_work_steal(int i,
5102 CMSParDrainMarkingStackClosure* drain,
5103 CMSParKeepAliveClosure* keep_alive,
5104 int* seed) {
5105 OopTaskQueue* work_q = work_queue(i);
5106 NOT_PRODUCT(int num_steals = 0;)
5107 oop obj_to_scan;
5108
5109 while (true) {
5110 // Completely finish any left over work from (an) earlier round(s)
5111 drain->trim_queue(0);
5112 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
5113 (size_t)ParGCDesiredObjsFromOverflowList);
5114 // Now check if there's any work in the overflow list
5115 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
5116 // only affects the number of attempts made to get work from the
5117 // overflow list and does not affect the number of workers. Just
5118 // pass ParallelGCThreads so this behavior is unchanged.
5119 if (_collector->par_take_from_overflow_list(num_from_overflow_list,
5120 work_q,
5121 ParallelGCThreads)) {
5122 // Found something in global overflow list;
5123 // not yet ready to go stealing work from others.
5124 // We'd like to assert(work_q->size() != 0, ...)
5125 // because we just took work from the overflow list,
5126 // but of course we can't, since all of that might have
5127 // been already stolen from us.
5128 continue;
5129 }
5130 // Verify that we have no work before we resort to stealing
5131 assert(work_q->size() == 0, "Have work, shouldn't steal");
5132 // Try to steal from other queues that have work
5133 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
5134 NOT_PRODUCT(num_steals++;)
5135 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!");
5136 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
5137 // Do scanning work
5138 obj_to_scan->oop_iterate(keep_alive);
5139 // Loop around, finish this work, and try to steal some more
5140 } else if (terminator()->offer_termination()) {
5141 break; // nirvana from the infinite cycle
5142 }
5143 }
5144 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals);
5145 }
5146
5147 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
5148 {
5149 CMSHeap* heap = CMSHeap::heap();
5150 WorkGang* workers = heap->workers();
5151 assert(workers != NULL, "Need parallel worker threads.");
5152 CMSRefProcTaskProxy rp_task(task, &_collector,
5153 _collector.ref_processor()->span(),
5154 _collector.markBitMap(),
5155 workers, _collector.task_queues());
5156 workers->run_task(&rp_task);
5157 }
5158
5159 void CMSRefProcTaskExecutor::execute(EnqueueTask& task)
5160 {
5161
5162 CMSHeap* heap = CMSHeap::heap();
5163 WorkGang* workers = heap->workers();
5164 assert(workers != NULL, "Need parallel worker threads.");
5165 CMSRefEnqueueTaskProxy enq_task(task);
5166 workers->run_task(&enq_task);
5167 }
5168
5169 void CMSCollector::refProcessingWork() {
5170 ResourceMark rm;
5171 HandleMark hm;
5172
5173 ReferenceProcessor* rp = ref_processor();
5174 assert(rp->span().equals(_span), "Spans should be equal");
5175 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete");
5176 // Process weak references.
5177 rp->setup_policy(false);
5178 verify_work_stacks_empty();
5179
5180 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q());
5181 {
5182 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm);
5183
5184 // Setup keep_alive and complete closures.
5185 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap,
5186 &_markStack, false /* !preclean */);
5187 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this,
5188 _span, &_markBitMap, &_markStack,
5189 &cmsKeepAliveClosure, false /* !preclean */);
5190
5191 ReferenceProcessorStats stats;
5192 if (rp->processing_is_mt()) {
5193 // Set the degree of MT here. If the discovery is done MT, there
5194 // may have been a different number of threads doing the discovery
5195 // and a different number of discovered lists may have Ref objects.
5196 // That is OK as long as the Reference lists are balanced (see
5197 // balance_all_queues() and balance_queues()).
5198 CMSHeap* heap = CMSHeap::heap();
5199 uint active_workers = ParallelGCThreads;
5200 WorkGang* workers = heap->workers();
5201 if (workers != NULL) {
5202 active_workers = workers->active_workers();
5203 // The expectation is that active_workers will have already
5204 // been set to a reasonable value. If it has not been set,
5205 // investigate.
5206 assert(active_workers > 0, "Should have been set during scavenge");
5207 }
5208 rp->set_active_mt_degree(active_workers);
5209 CMSRefProcTaskExecutor task_executor(*this);
5210 stats = rp->process_discovered_references(&_is_alive_closure,
5211 &cmsKeepAliveClosure,
5212 &cmsDrainMarkingStackClosure,
5213 &task_executor,
5214 &pt);
5215 } else {
5216 stats = rp->process_discovered_references(&_is_alive_closure,
5217 &cmsKeepAliveClosure,
5218 &cmsDrainMarkingStackClosure,
5219 NULL,
5220 &pt);
5221 }
5222 _gc_tracer_cm->report_gc_reference_stats(stats);
5223 pt.print_all_references();
5224 }
5225
5226 // This is the point where the entire marking should have completed.
5227 verify_work_stacks_empty();
5228
5229 {
5230 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm);
5231 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl);
5232 }
5233
5234 if (should_unload_classes()) {
5235 {
5236 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm);
5237
5238 // Unload classes and purge the SystemDictionary.
5239 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm);
5240
5241 // Unload nmethods.
5242 CodeCache::do_unloading(&_is_alive_closure, purged_class);
5243
5244 // Prune dead klasses from subklass/sibling/implementor lists.
5245 Klass::clean_weak_klass_links(&_is_alive_closure);
5246 }
5247
5248 {
5249 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm);
5250 // Clean up unreferenced symbols in symbol table.
5251 SymbolTable::unlink();
5252 }
5253
5254 {
5255 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm);
5256 // Delete entries for dead interned strings.
5257 StringTable::unlink(&_is_alive_closure);
5258 }
5259 }
5260
5261 // Restore any preserved marks as a result of mark stack or
5262 // work queue overflow
5263 restore_preserved_marks_if_any(); // done single-threaded for now
5264
5265 rp->set_enqueuing_is_done(true);
5266 if (rp->processing_is_mt()) {
5267 rp->balance_all_queues();
5268 CMSRefProcTaskExecutor task_executor(*this);
5269 rp->enqueue_discovered_references(&task_executor, &pt);
5270 } else {
5271 rp->enqueue_discovered_references(NULL, &pt);
5272 }
5273 rp->verify_no_references_recorded();
5274 pt.print_enqueue_phase();
5275 assert(!rp->discovery_enabled(), "should have been disabled");
5276 }
5277
5278 #ifndef PRODUCT
5279 void CMSCollector::check_correct_thread_executing() {
5280 Thread* t = Thread::current();
5281 // Only the VM thread or the CMS thread should be here.
5282 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(),
5283 "Unexpected thread type");
5284 // If this is the vm thread, the foreground process
5285 // should not be waiting. Note that _foregroundGCIsActive is
5286 // true while the foreground collector is waiting.
5287 if (_foregroundGCShouldWait) {
5288 // We cannot be the VM thread
5289 assert(t->is_ConcurrentGC_thread(),
5290 "Should be CMS thread");
5291 } else {
5292 // We can be the CMS thread only if we are in a stop-world
5293 // phase of CMS collection.
5294 if (t->is_ConcurrentGC_thread()) {
5295 assert(_collectorState == InitialMarking ||
5296 _collectorState == FinalMarking,
5297 "Should be a stop-world phase");
5298 // The CMS thread should be holding the CMS_token.
5299 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5300 "Potential interference with concurrently "
5301 "executing VM thread");
5302 }
5303 }
5304 }
5305 #endif
5306
5307 void CMSCollector::sweep() {
5308 assert(_collectorState == Sweeping, "just checking");
5309 check_correct_thread_executing();
5310 verify_work_stacks_empty();
5311 verify_overflow_empty();
5312 increment_sweep_count();
5313 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause());
5314
5315 _inter_sweep_timer.stop();
5316 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds());
5317
5318 assert(!_intra_sweep_timer.is_active(), "Should not be active");
5319 _intra_sweep_timer.reset();
5320 _intra_sweep_timer.start();
5321 {
5322 GCTraceCPUTime tcpu;
5323 CMSPhaseAccounting pa(this, "Concurrent Sweep");
5324 // First sweep the old gen
5325 {
5326 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(),
5327 bitMapLock());
5328 sweepWork(_cmsGen);
5329 }
5330
5331 // Update Universe::_heap_*_at_gc figures.
5332 // We need all the free list locks to make the abstract state
5333 // transition from Sweeping to Resetting. See detailed note
5334 // further below.
5335 {
5336 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock());
5337 // Update heap occupancy information which is used as
5338 // input to soft ref clearing policy at the next gc.
5339 Universe::update_heap_info_at_gc();
5340 _collectorState = Resizing;
5341 }
5342 }
5343 verify_work_stacks_empty();
5344 verify_overflow_empty();
5345
5346 if (should_unload_classes()) {
5347 // Delay purge to the beginning of the next safepoint. Metaspace::contains
5348 // requires that the virtual spaces are stable and not deleted.
5349 ClassLoaderDataGraph::set_should_purge(true);
5350 }
5351
5352 _intra_sweep_timer.stop();
5353 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds());
5354
5355 _inter_sweep_timer.reset();
5356 _inter_sweep_timer.start();
5357
5358 // We need to use a monotonically non-decreasing time in ms
5359 // or we will see time-warp warnings and os::javaTimeMillis()
5360 // does not guarantee monotonicity.
5361 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
5362 update_time_of_last_gc(now);
5363
5364 // NOTE on abstract state transitions:
5365 // Mutators allocate-live and/or mark the mod-union table dirty
5366 // based on the state of the collection. The former is done in
5367 // the interval [Marking, Sweeping] and the latter in the interval
5368 // [Marking, Sweeping). Thus the transitions into the Marking state
5369 // and out of the Sweeping state must be synchronously visible
5370 // globally to the mutators.
5371 // The transition into the Marking state happens with the world
5372 // stopped so the mutators will globally see it. Sweeping is
5373 // done asynchronously by the background collector so the transition
5374 // from the Sweeping state to the Resizing state must be done
5375 // under the freelistLock (as is the check for whether to
5376 // allocate-live and whether to dirty the mod-union table).
5377 assert(_collectorState == Resizing, "Change of collector state to"
5378 " Resizing must be done under the freelistLocks (plural)");
5379
5380 // Now that sweeping has been completed, we clear
5381 // the incremental_collection_failed flag,
5382 // thus inviting a younger gen collection to promote into
5383 // this generation. If such a promotion may still fail,
5384 // the flag will be set again when a young collection is
5385 // attempted.
5386 CMSHeap* heap = CMSHeap::heap();
5387 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up
5388 heap->update_full_collections_completed(_collection_count_start);
5389 }
5390
5391 // FIX ME!!! Looks like this belongs in CFLSpace, with
5392 // CMSGen merely delegating to it.
5393 void ConcurrentMarkSweepGeneration::setNearLargestChunk() {
5394 double nearLargestPercent = FLSLargestBlockCoalesceProximity;
5395 HeapWord* minAddr = _cmsSpace->bottom();
5396 HeapWord* largestAddr =
5397 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict();
5398 if (largestAddr == NULL) {
5399 // The dictionary appears to be empty. In this case
5400 // try to coalesce at the end of the heap.
5401 largestAddr = _cmsSpace->end();
5402 }
5403 size_t largestOffset = pointer_delta(largestAddr, minAddr);
5404 size_t nearLargestOffset =
5405 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize;
5406 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT,
5407 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset));
5408 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset);
5409 }
5410
5411 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) {
5412 return addr >= _cmsSpace->nearLargestChunk();
5413 }
5414
5415 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() {
5416 return _cmsSpace->find_chunk_at_end();
5417 }
5418
5419 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation,
5420 bool full) {
5421 // If the young generation has been collected, gather any statistics
5422 // that are of interest at this point.
5423 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation);
5424 if (!full && current_is_young) {
5425 // Gather statistics on the young generation collection.
5426 collector()->stats().record_gc0_end(used());
5427 }
5428 }
5429
5430 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) {
5431 // We iterate over the space(s) underlying this generation,
5432 // checking the mark bit map to see if the bits corresponding
5433 // to specific blocks are marked or not. Blocks that are
5434 // marked are live and are not swept up. All remaining blocks
5435 // are swept up, with coalescing on-the-fly as we sweep up
5436 // contiguous free and/or garbage blocks:
5437 // We need to ensure that the sweeper synchronizes with allocators
5438 // and stop-the-world collectors. In particular, the following
5439 // locks are used:
5440 // . CMS token: if this is held, a stop the world collection cannot occur
5441 // . freelistLock: if this is held no allocation can occur from this
5442 // generation by another thread
5443 // . bitMapLock: if this is held, no other thread can access or update
5444 //
5445
5446 // Note that we need to hold the freelistLock if we use
5447 // block iterate below; else the iterator might go awry if
5448 // a mutator (or promotion) causes block contents to change
5449 // (for instance if the allocator divvies up a block).
5450 // If we hold the free list lock, for all practical purposes
5451 // young generation GC's can't occur (they'll usually need to
5452 // promote), so we might as well prevent all young generation
5453 // GC's while we do a sweeping step. For the same reason, we might
5454 // as well take the bit map lock for the entire duration
5455
5456 // check that we hold the requisite locks
5457 assert(have_cms_token(), "Should hold cms token");
5458 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep");
5459 assert_lock_strong(old_gen->freelistLock());
5460 assert_lock_strong(bitMapLock());
5461
5462 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context");
5463 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context");
5464 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()),
5465 _inter_sweep_estimate.padded_average(),
5466 _intra_sweep_estimate.padded_average());
5467 old_gen->setNearLargestChunk();
5468
5469 {
5470 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield);
5471 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure);
5472 // We need to free-up/coalesce garbage/blocks from a
5473 // co-terminal free run. This is done in the SweepClosure
5474 // destructor; so, do not remove this scope, else the
5475 // end-of-sweep-census below will be off by a little bit.
5476 }
5477 old_gen->cmsSpace()->sweep_completed();
5478 old_gen->cmsSpace()->endSweepFLCensus(sweep_count());
5479 if (should_unload_classes()) { // unloaded classes this cycle,
5480 _concurrent_cycles_since_last_unload = 0; // ... reset count
5481 } else { // did not unload classes,
5482 _concurrent_cycles_since_last_unload++; // ... increment count
5483 }
5484 }
5485
5486 // Reset CMS data structures (for now just the marking bit map)
5487 // preparatory for the next cycle.
5488 void CMSCollector::reset_concurrent() {
5489 CMSTokenSyncWithLocks ts(true, bitMapLock());
5490
5491 // If the state is not "Resetting", the foreground thread
5492 // has done a collection and the resetting.
5493 if (_collectorState != Resetting) {
5494 assert(_collectorState == Idling, "The state should only change"
5495 " because the foreground collector has finished the collection");
5496 return;
5497 }
5498
5499 {
5500 // Clear the mark bitmap (no grey objects to start with)
5501 // for the next cycle.
5502 GCTraceCPUTime tcpu;
5503 CMSPhaseAccounting cmspa(this, "Concurrent Reset");
5504
5505 HeapWord* curAddr = _markBitMap.startWord();
5506 while (curAddr < _markBitMap.endWord()) {
5507 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr);
5508 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining));
5509 _markBitMap.clear_large_range(chunk);
5510 if (ConcurrentMarkSweepThread::should_yield() &&
5511 !foregroundGCIsActive() &&
5512 CMSYield) {
5513 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5514 "CMS thread should hold CMS token");
5515 assert_lock_strong(bitMapLock());
5516 bitMapLock()->unlock();
5517 ConcurrentMarkSweepThread::desynchronize(true);
5518 stopTimer();
5519 incrementYields();
5520
5521 // See the comment in coordinator_yield()
5522 for (unsigned i = 0; i < CMSYieldSleepCount &&
5523 ConcurrentMarkSweepThread::should_yield() &&
5524 !CMSCollector::foregroundGCIsActive(); ++i) {
5525 os::sleep(Thread::current(), 1, false);
5526 }
5527
5528 ConcurrentMarkSweepThread::synchronize(true);
5529 bitMapLock()->lock_without_safepoint_check();
5530 startTimer();
5531 }
5532 curAddr = chunk.end();
5533 }
5534 // A successful mostly concurrent collection has been done.
5535 // Because only the full (i.e., concurrent mode failure) collections
5536 // are being measured for gc overhead limits, clean the "near" flag
5537 // and count.
5538 size_policy()->reset_gc_overhead_limit_count();
5539 _collectorState = Idling;
5540 }
5541
5542 register_gc_end();
5543 }
5544
5545 // Same as above but for STW paths
5546 void CMSCollector::reset_stw() {
5547 // already have the lock
5548 assert(_collectorState == Resetting, "just checking");
5549 assert_lock_strong(bitMapLock());
5550 GCIdMark gc_id_mark(_cmsThread->gc_id());
5551 _markBitMap.clear_all();
5552 _collectorState = Idling;
5553 register_gc_end();
5554 }
5555
5556 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) {
5557 GCTraceCPUTime tcpu;
5558 TraceCollectorStats tcs_cgc(cgc_counters());
5559
5560 switch (op) {
5561 case CMS_op_checkpointRootsInitial: {
5562 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true);
5563 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5564 checkpointRootsInitial();
5565 break;
5566 }
5567 case CMS_op_checkpointRootsFinal: {
5568 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true);
5569 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT);
5570 checkpointRootsFinal();
5571 break;
5572 }
5573 default:
5574 fatal("No such CMS_op");
5575 }
5576 }
5577
5578 #ifndef PRODUCT
5579 size_t const CMSCollector::skip_header_HeapWords() {
5580 return FreeChunk::header_size();
5581 }
5582
5583 // Try and collect here conditions that should hold when
5584 // CMS thread is exiting. The idea is that the foreground GC
5585 // thread should not be blocked if it wants to terminate
5586 // the CMS thread and yet continue to run the VM for a while
5587 // after that.
5588 void CMSCollector::verify_ok_to_terminate() const {
5589 assert(Thread::current()->is_ConcurrentGC_thread(),
5590 "should be called by CMS thread");
5591 assert(!_foregroundGCShouldWait, "should be false");
5592 // We could check here that all the various low-level locks
5593 // are not held by the CMS thread, but that is overkill; see
5594 // also CMSThread::verify_ok_to_terminate() where the CGC_lock
5595 // is checked.
5596 }
5597 #endif
5598
5599 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const {
5600 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1),
5601 "missing Printezis mark?");
5602 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5603 size_t size = pointer_delta(nextOneAddr + 1, addr);
5604 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5605 "alignment problem");
5606 assert(size >= 3, "Necessary for Printezis marks to work");
5607 return size;
5608 }
5609
5610 // A variant of the above (block_size_using_printezis_bits()) except
5611 // that we return 0 if the P-bits are not yet set.
5612 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const {
5613 if (_markBitMap.isMarked(addr + 1)) {
5614 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects");
5615 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2);
5616 size_t size = pointer_delta(nextOneAddr + 1, addr);
5617 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
5618 "alignment problem");
5619 assert(size >= 3, "Necessary for Printezis marks to work");
5620 return size;
5621 }
5622 return 0;
5623 }
5624
5625 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const {
5626 size_t sz = 0;
5627 oop p = (oop)addr;
5628 if (p->klass_or_null_acquire() != NULL) {
5629 sz = CompactibleFreeListSpace::adjustObjectSize(p->size());
5630 } else {
5631 sz = block_size_using_printezis_bits(addr);
5632 }
5633 assert(sz > 0, "size must be nonzero");
5634 HeapWord* next_block = addr + sz;
5635 HeapWord* next_card = align_up(next_block, CardTable::card_size);
5636 assert(align_down((uintptr_t)addr, CardTable::card_size) <
5637 align_down((uintptr_t)next_card, CardTable::card_size),
5638 "must be different cards");
5639 return next_card;
5640 }
5641
5642
5643 // CMS Bit Map Wrapper /////////////////////////////////////////
5644
5645 // Construct a CMS bit map infrastructure, but don't create the
5646 // bit vector itself. That is done by a separate call CMSBitMap::allocate()
5647 // further below.
5648 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name):
5649 _bm(),
5650 _shifter(shifter),
5651 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true,
5652 Monitor::_safepoint_check_sometimes) : NULL)
5653 {
5654 _bmStartWord = 0;
5655 _bmWordSize = 0;
5656 }
5657
5658 bool CMSBitMap::allocate(MemRegion mr) {
5659 _bmStartWord = mr.start();
5660 _bmWordSize = mr.word_size();
5661 ReservedSpace brs(ReservedSpace::allocation_align_size_up(
5662 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
5663 if (!brs.is_reserved()) {
5664 log_warning(gc)("CMS bit map allocation failure");
5665 return false;
5666 }
5667 // For now we'll just commit all of the bit map up front.
5668 // Later on we'll try to be more parsimonious with swap.
5669 if (!_virtual_space.initialize(brs, brs.size())) {
5670 log_warning(gc)("CMS bit map backing store failure");
5671 return false;
5672 }
5673 assert(_virtual_space.committed_size() == brs.size(),
5674 "didn't reserve backing store for all of CMS bit map?");
5675 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
5676 _bmWordSize, "inconsistency in bit map sizing");
5677 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter);
5678
5679 // bm.clear(); // can we rely on getting zero'd memory? verify below
5680 assert(isAllClear(),
5681 "Expected zero'd memory from ReservedSpace constructor");
5682 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()),
5683 "consistency check");
5684 return true;
5685 }
5686
5687 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) {
5688 HeapWord *next_addr, *end_addr, *last_addr;
5689 assert_locked();
5690 assert(covers(mr), "out-of-range error");
5691 // XXX assert that start and end are appropriately aligned
5692 for (next_addr = mr.start(), end_addr = mr.end();
5693 next_addr < end_addr; next_addr = last_addr) {
5694 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr);
5695 last_addr = dirty_region.end();
5696 if (!dirty_region.is_empty()) {
5697 cl->do_MemRegion(dirty_region);
5698 } else {
5699 assert(last_addr == end_addr, "program logic");
5700 return;
5701 }
5702 }
5703 }
5704
5705 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const {
5706 _bm.print_on_error(st, prefix);
5707 }
5708
5709 #ifndef PRODUCT
5710 void CMSBitMap::assert_locked() const {
5711 CMSLockVerifier::assert_locked(lock());
5712 }
5713
5714 bool CMSBitMap::covers(MemRegion mr) const {
5715 // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
5716 assert((size_t)_bm.size() == (_bmWordSize >> _shifter),
5717 "size inconsistency");
5718 return (mr.start() >= _bmStartWord) &&
5719 (mr.end() <= endWord());
5720 }
5721
5722 bool CMSBitMap::covers(HeapWord* start, size_t size) const {
5723 return (start >= _bmStartWord && (start + size) <= endWord());
5724 }
5725
5726 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) {
5727 // verify that there are no 1 bits in the interval [left, right)
5728 FalseBitMapClosure falseBitMapClosure;
5729 iterate(&falseBitMapClosure, left, right);
5730 }
5731
5732 void CMSBitMap::region_invariant(MemRegion mr)
5733 {
5734 assert_locked();
5735 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
5736 assert(!mr.is_empty(), "unexpected empty region");
5737 assert(covers(mr), "mr should be covered by bit map");
5738 // convert address range into offset range
5739 size_t start_ofs = heapWordToOffset(mr.start());
5740 // Make sure that end() is appropriately aligned
5741 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))),
5742 "Misaligned mr.end()");
5743 size_t end_ofs = heapWordToOffset(mr.end());
5744 assert(end_ofs > start_ofs, "Should mark at least one bit");
5745 }
5746
5747 #endif
5748
5749 bool CMSMarkStack::allocate(size_t size) {
5750 // allocate a stack of the requisite depth
5751 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5752 size * sizeof(oop)));
5753 if (!rs.is_reserved()) {
5754 log_warning(gc)("CMSMarkStack allocation failure");
5755 return false;
5756 }
5757 if (!_virtual_space.initialize(rs, rs.size())) {
5758 log_warning(gc)("CMSMarkStack backing store failure");
5759 return false;
5760 }
5761 assert(_virtual_space.committed_size() == rs.size(),
5762 "didn't reserve backing store for all of CMS stack?");
5763 _base = (oop*)(_virtual_space.low());
5764 _index = 0;
5765 _capacity = size;
5766 NOT_PRODUCT(_max_depth = 0);
5767 return true;
5768 }
5769
5770 // XXX FIX ME !!! In the MT case we come in here holding a
5771 // leaf lock. For printing we need to take a further lock
5772 // which has lower rank. We need to recalibrate the two
5773 // lock-ranks involved in order to be able to print the
5774 // messages below. (Or defer the printing to the caller.
5775 // For now we take the expedient path of just disabling the
5776 // messages for the problematic case.)
5777 void CMSMarkStack::expand() {
5778 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted");
5779 if (_capacity == MarkStackSizeMax) {
5780 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) {
5781 // We print a warning message only once per CMS cycle.
5782 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit");
5783 }
5784 return;
5785 }
5786 // Double capacity if possible
5787 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax);
5788 // Do not give up existing stack until we have managed to
5789 // get the double capacity that we desired.
5790 ReservedSpace rs(ReservedSpace::allocation_align_size_up(
5791 new_capacity * sizeof(oop)));
5792 if (rs.is_reserved()) {
5793 // Release the backing store associated with old stack
5794 _virtual_space.release();
5795 // Reinitialize virtual space for new stack
5796 if (!_virtual_space.initialize(rs, rs.size())) {
5797 fatal("Not enough swap for expanded marking stack");
5798 }
5799 _base = (oop*)(_virtual_space.low());
5800 _index = 0;
5801 _capacity = new_capacity;
5802 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) {
5803 // Failed to double capacity, continue;
5804 // we print a detail message only once per CMS cycle.
5805 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
5806 _capacity / K, new_capacity / K);
5807 }
5808 }
5809
5810
5811 // Closures
5812 // XXX: there seems to be a lot of code duplication here;
5813 // should refactor and consolidate common code.
5814
5815 // This closure is used to mark refs into the CMS generation in
5816 // the CMS bit map. Called at the first checkpoint. This closure
5817 // assumes that we do not need to re-mark dirty cards; if the CMS
5818 // generation on which this is used is not an oldest
5819 // generation then this will lose younger_gen cards!
5820
5821 MarkRefsIntoClosure::MarkRefsIntoClosure(
5822 MemRegion span, CMSBitMap* bitMap):
5823 _span(span),
5824 _bitMap(bitMap)
5825 {
5826 assert(ref_discoverer() == NULL, "deliberately left NULL");
5827 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5828 }
5829
5830 void MarkRefsIntoClosure::do_oop(oop obj) {
5831 // if p points into _span, then mark corresponding bit in _markBitMap
5832 assert(oopDesc::is_oop(obj), "expected an oop");
5833 HeapWord* addr = (HeapWord*)obj;
5834 if (_span.contains(addr)) {
5835 // this should be made more efficient
5836 _bitMap->mark(addr);
5837 }
5838 }
5839
5840 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5841 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); }
5842
5843 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure(
5844 MemRegion span, CMSBitMap* bitMap):
5845 _span(span),
5846 _bitMap(bitMap)
5847 {
5848 assert(ref_discoverer() == NULL, "deliberately left NULL");
5849 assert(_bitMap->covers(_span), "_bitMap/_span mismatch");
5850 }
5851
5852 void ParMarkRefsIntoClosure::do_oop(oop obj) {
5853 // if p points into _span, then mark corresponding bit in _markBitMap
5854 assert(oopDesc::is_oop(obj), "expected an oop");
5855 HeapWord* addr = (HeapWord*)obj;
5856 if (_span.contains(addr)) {
5857 // this should be made more efficient
5858 _bitMap->par_mark(addr);
5859 }
5860 }
5861
5862 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5863 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); }
5864
5865 // A variant of the above, used for CMS marking verification.
5866 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure(
5867 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm):
5868 _span(span),
5869 _verification_bm(verification_bm),
5870 _cms_bm(cms_bm)
5871 {
5872 assert(ref_discoverer() == NULL, "deliberately left NULL");
5873 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch");
5874 }
5875
5876 void MarkRefsIntoVerifyClosure::do_oop(oop obj) {
5877 // if p points into _span, then mark corresponding bit in _markBitMap
5878 assert(oopDesc::is_oop(obj), "expected an oop");
5879 HeapWord* addr = (HeapWord*)obj;
5880 if (_span.contains(addr)) {
5881 _verification_bm->mark(addr);
5882 if (!_cms_bm->isMarked(addr)) {
5883 Log(gc, verify) log;
5884 ResourceMark rm;
5885 LogStream ls(log.error());
5886 oop(addr)->print_on(&ls);
5887 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
5888 fatal("... aborting");
5889 }
5890 }
5891 }
5892
5893 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5894 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); }
5895
5896 //////////////////////////////////////////////////
5897 // MarkRefsIntoAndScanClosure
5898 //////////////////////////////////////////////////
5899
5900 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span,
5901 ReferenceDiscoverer* rd,
5902 CMSBitMap* bit_map,
5903 CMSBitMap* mod_union_table,
5904 CMSMarkStack* mark_stack,
5905 CMSCollector* collector,
5906 bool should_yield,
5907 bool concurrent_precleaning):
5908 _collector(collector),
5909 _span(span),
5910 _bit_map(bit_map),
5911 _mark_stack(mark_stack),
5912 _pushAndMarkClosure(collector, span, rd, bit_map, mod_union_table,
5913 mark_stack, concurrent_precleaning),
5914 _yield(should_yield),
5915 _concurrent_precleaning(concurrent_precleaning),
5916 _freelistLock(NULL)
5917 {
5918 // FIXME: Should initialize in base class constructor.
5919 assert(rd != NULL, "ref_discoverer shouldn't be NULL");
5920 set_ref_discoverer_internal(rd);
5921 }
5922
5923 // This closure is used to mark refs into the CMS generation at the
5924 // second (final) checkpoint, and to scan and transitively follow
5925 // the unmarked oops. It is also used during the concurrent precleaning
5926 // phase while scanning objects on dirty cards in the CMS generation.
5927 // The marks are made in the marking bit map and the marking stack is
5928 // used for keeping the (newly) grey objects during the scan.
5929 // The parallel version (Par_...) appears further below.
5930 void MarkRefsIntoAndScanClosure::do_oop(oop obj) {
5931 if (obj != NULL) {
5932 assert(oopDesc::is_oop(obj), "expected an oop");
5933 HeapWord* addr = (HeapWord*)obj;
5934 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
5935 assert(_collector->overflow_list_is_empty(),
5936 "overflow list should be empty");
5937 if (_span.contains(addr) &&
5938 !_bit_map->isMarked(addr)) {
5939 // mark bit map (object is now grey)
5940 _bit_map->mark(addr);
5941 // push on marking stack (stack should be empty), and drain the
5942 // stack by applying this closure to the oops in the oops popped
5943 // from the stack (i.e. blacken the grey objects)
5944 bool res = _mark_stack->push(obj);
5945 assert(res, "Should have space to push on empty stack");
5946 do {
5947 oop new_oop = _mark_stack->pop();
5948 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
5949 assert(_bit_map->isMarked((HeapWord*)new_oop),
5950 "only grey objects on this stack");
5951 // iterate over the oops in this oop, marking and pushing
5952 // the ones in CMS heap (i.e. in _span).
5953 new_oop->oop_iterate(&_pushAndMarkClosure);
5954 // check if it's time to yield
5955 do_yield_check();
5956 } while (!_mark_stack->isEmpty() ||
5957 (!_concurrent_precleaning && take_from_overflow_list()));
5958 // if marking stack is empty, and we are not doing this
5959 // during precleaning, then check the overflow list
5960 }
5961 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
5962 assert(_collector->overflow_list_is_empty(),
5963 "overflow list was drained above");
5964
5965 assert(_collector->no_preserved_marks(),
5966 "All preserved marks should have been restored above");
5967 }
5968 }
5969
5970 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5971 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); }
5972
5973 void MarkRefsIntoAndScanClosure::do_yield_work() {
5974 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
5975 "CMS thread should hold CMS token");
5976 assert_lock_strong(_freelistLock);
5977 assert_lock_strong(_bit_map->lock());
5978 // relinquish the free_list_lock and bitMaplock()
5979 _bit_map->lock()->unlock();
5980 _freelistLock->unlock();
5981 ConcurrentMarkSweepThread::desynchronize(true);
5982 _collector->stopTimer();
5983 _collector->incrementYields();
5984
5985 // See the comment in coordinator_yield()
5986 for (unsigned i = 0;
5987 i < CMSYieldSleepCount &&
5988 ConcurrentMarkSweepThread::should_yield() &&
5989 !CMSCollector::foregroundGCIsActive();
5990 ++i) {
5991 os::sleep(Thread::current(), 1, false);
5992 }
5993
5994 ConcurrentMarkSweepThread::synchronize(true);
5995 _freelistLock->lock_without_safepoint_check();
5996 _bit_map->lock()->lock_without_safepoint_check();
5997 _collector->startTimer();
5998 }
5999
6000 ///////////////////////////////////////////////////////////
6001 // ParMarkRefsIntoAndScanClosure: a parallel version of
6002 // MarkRefsIntoAndScanClosure
6003 ///////////////////////////////////////////////////////////
6004 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure(
6005 CMSCollector* collector, MemRegion span, ReferenceDiscoverer* rd,
6006 CMSBitMap* bit_map, OopTaskQueue* work_queue):
6007 _span(span),
6008 _bit_map(bit_map),
6009 _work_queue(work_queue),
6010 _low_water_mark(MIN2((work_queue->max_elems()/4),
6011 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))),
6012 _parPushAndMarkClosure(collector, span, rd, bit_map, work_queue)
6013 {
6014 // FIXME: Should initialize in base class constructor.
6015 assert(rd != NULL, "ref_discoverer shouldn't be NULL");
6016 set_ref_discoverer_internal(rd);
6017 }
6018
6019 // This closure is used to mark refs into the CMS generation at the
6020 // second (final) checkpoint, and to scan and transitively follow
6021 // the unmarked oops. The marks are made in the marking bit map and
6022 // the work_queue is used for keeping the (newly) grey objects during
6023 // the scan phase whence they are also available for stealing by parallel
6024 // threads. Since the marking bit map is shared, updates are
6025 // synchronized (via CAS).
6026 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) {
6027 if (obj != NULL) {
6028 // Ignore mark word because this could be an already marked oop
6029 // that may be chained at the end of the overflow list.
6030 assert(oopDesc::is_oop(obj, true), "expected an oop");
6031 HeapWord* addr = (HeapWord*)obj;
6032 if (_span.contains(addr) &&
6033 !_bit_map->isMarked(addr)) {
6034 // mark bit map (object will become grey):
6035 // It is possible for several threads to be
6036 // trying to "claim" this object concurrently;
6037 // the unique thread that succeeds in marking the
6038 // object first will do the subsequent push on
6039 // to the work queue (or overflow list).
6040 if (_bit_map->par_mark(addr)) {
6041 // push on work_queue (which may not be empty), and trim the
6042 // queue to an appropriate length by applying this closure to
6043 // the oops in the oops popped from the stack (i.e. blacken the
6044 // grey objects)
6045 bool res = _work_queue->push(obj);
6046 assert(res, "Low water mark should be less than capacity?");
6047 trim_queue(_low_water_mark);
6048 } // Else, another thread claimed the object
6049 }
6050 }
6051 }
6052
6053 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6054 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); }
6055
6056 // This closure is used to rescan the marked objects on the dirty cards
6057 // in the mod union table and the card table proper.
6058 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m(
6059 oop p, MemRegion mr) {
6060
6061 size_t size = 0;
6062 HeapWord* addr = (HeapWord*)p;
6063 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6064 assert(_span.contains(addr), "we are scanning the CMS generation");
6065 // check if it's time to yield
6066 if (do_yield_check()) {
6067 // We yielded for some foreground stop-world work,
6068 // and we have been asked to abort this ongoing preclean cycle.
6069 return 0;
6070 }
6071 if (_bitMap->isMarked(addr)) {
6072 // it's marked; is it potentially uninitialized?
6073 if (p->klass_or_null_acquire() != NULL) {
6074 // an initialized object; ignore mark word in verification below
6075 // since we are running concurrent with mutators
6076 assert(oopDesc::is_oop(p, true), "should be an oop");
6077 if (p->is_objArray()) {
6078 // objArrays are precisely marked; restrict scanning
6079 // to dirty cards only.
6080 size = CompactibleFreeListSpace::adjustObjectSize(
6081 p->oop_iterate_size(_scanningClosure, mr));
6082 } else {
6083 // A non-array may have been imprecisely marked; we need
6084 // to scan object in its entirety.
6085 size = CompactibleFreeListSpace::adjustObjectSize(
6086 p->oop_iterate_size(_scanningClosure));
6087 }
6088 #ifdef ASSERT
6089 size_t direct_size =
6090 CompactibleFreeListSpace::adjustObjectSize(p->size());
6091 assert(size == direct_size, "Inconsistency in size");
6092 assert(size >= 3, "Necessary for Printezis marks to work");
6093 HeapWord* start_pbit = addr + 1;
6094 HeapWord* end_pbit = addr + size - 1;
6095 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit),
6096 "inconsistent Printezis mark");
6097 // Verify inner mark bits (between Printezis bits) are clear,
6098 // but don't repeat if there are multiple dirty regions for
6099 // the same object, to avoid potential O(N^2) performance.
6100 if (addr != _last_scanned_object) {
6101 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit);
6102 _last_scanned_object = addr;
6103 }
6104 #endif // ASSERT
6105 } else {
6106 // An uninitialized object.
6107 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?");
6108 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
6109 size = pointer_delta(nextOneAddr + 1, addr);
6110 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
6111 "alignment problem");
6112 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass()
6113 // will dirty the card when the klass pointer is installed in the
6114 // object (signaling the completion of initialization).
6115 }
6116 } else {
6117 // Either a not yet marked object or an uninitialized object
6118 if (p->klass_or_null_acquire() == NULL) {
6119 // An uninitialized object, skip to the next card, since
6120 // we may not be able to read its P-bits yet.
6121 assert(size == 0, "Initial value");
6122 } else {
6123 // An object not (yet) reached by marking: we merely need to
6124 // compute its size so as to go look at the next block.
6125 assert(oopDesc::is_oop(p, true), "should be an oop");
6126 size = CompactibleFreeListSpace::adjustObjectSize(p->size());
6127 }
6128 }
6129 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6130 return size;
6131 }
6132
6133 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() {
6134 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6135 "CMS thread should hold CMS token");
6136 assert_lock_strong(_freelistLock);
6137 assert_lock_strong(_bitMap->lock());
6138 // relinquish the free_list_lock and bitMaplock()
6139 _bitMap->lock()->unlock();
6140 _freelistLock->unlock();
6141 ConcurrentMarkSweepThread::desynchronize(true);
6142 _collector->stopTimer();
6143 _collector->incrementYields();
6144
6145 // See the comment in coordinator_yield()
6146 for (unsigned i = 0; i < CMSYieldSleepCount &&
6147 ConcurrentMarkSweepThread::should_yield() &&
6148 !CMSCollector::foregroundGCIsActive(); ++i) {
6149 os::sleep(Thread::current(), 1, false);
6150 }
6151
6152 ConcurrentMarkSweepThread::synchronize(true);
6153 _freelistLock->lock_without_safepoint_check();
6154 _bitMap->lock()->lock_without_safepoint_check();
6155 _collector->startTimer();
6156 }
6157
6158
6159 //////////////////////////////////////////////////////////////////
6160 // SurvivorSpacePrecleanClosure
6161 //////////////////////////////////////////////////////////////////
6162 // This (single-threaded) closure is used to preclean the oops in
6163 // the survivor spaces.
6164 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) {
6165
6166 HeapWord* addr = (HeapWord*)p;
6167 DEBUG_ONLY(_collector->verify_work_stacks_empty();)
6168 assert(!_span.contains(addr), "we are scanning the survivor spaces");
6169 assert(p->klass_or_null() != NULL, "object should be initialized");
6170 // an initialized object; ignore mark word in verification below
6171 // since we are running concurrent with mutators
6172 assert(oopDesc::is_oop(p, true), "should be an oop");
6173 // Note that we do not yield while we iterate over
6174 // the interior oops of p, pushing the relevant ones
6175 // on our marking stack.
6176 size_t size = p->oop_iterate_size(_scanning_closure);
6177 do_yield_check();
6178 // Observe that below, we do not abandon the preclean
6179 // phase as soon as we should; rather we empty the
6180 // marking stack before returning. This is to satisfy
6181 // some existing assertions. In general, it may be a
6182 // good idea to abort immediately and complete the marking
6183 // from the grey objects at a later time.
6184 while (!_mark_stack->isEmpty()) {
6185 oop new_oop = _mark_stack->pop();
6186 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
6187 assert(_bit_map->isMarked((HeapWord*)new_oop),
6188 "only grey objects on this stack");
6189 // iterate over the oops in this oop, marking and pushing
6190 // the ones in CMS heap (i.e. in _span).
6191 new_oop->oop_iterate(_scanning_closure);
6192 // check if it's time to yield
6193 do_yield_check();
6194 }
6195 unsigned int after_count =
6196 CMSHeap::heap()->total_collections();
6197 bool abort = (_before_count != after_count) ||
6198 _collector->should_abort_preclean();
6199 return abort ? 0 : size;
6200 }
6201
6202 void SurvivorSpacePrecleanClosure::do_yield_work() {
6203 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6204 "CMS thread should hold CMS token");
6205 assert_lock_strong(_bit_map->lock());
6206 // Relinquish the bit map lock
6207 _bit_map->lock()->unlock();
6208 ConcurrentMarkSweepThread::desynchronize(true);
6209 _collector->stopTimer();
6210 _collector->incrementYields();
6211
6212 // See the comment in coordinator_yield()
6213 for (unsigned i = 0; i < CMSYieldSleepCount &&
6214 ConcurrentMarkSweepThread::should_yield() &&
6215 !CMSCollector::foregroundGCIsActive(); ++i) {
6216 os::sleep(Thread::current(), 1, false);
6217 }
6218
6219 ConcurrentMarkSweepThread::synchronize(true);
6220 _bit_map->lock()->lock_without_safepoint_check();
6221 _collector->startTimer();
6222 }
6223
6224 // This closure is used to rescan the marked objects on the dirty cards
6225 // in the mod union table and the card table proper. In the parallel
6226 // case, although the bitMap is shared, we do a single read so the
6227 // isMarked() query is "safe".
6228 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) {
6229 // Ignore mark word because we are running concurrent with mutators
6230 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p));
6231 HeapWord* addr = (HeapWord*)p;
6232 assert(_span.contains(addr), "we are scanning the CMS generation");
6233 bool is_obj_array = false;
6234 #ifdef ASSERT
6235 if (!_parallel) {
6236 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)");
6237 assert(_collector->overflow_list_is_empty(),
6238 "overflow list should be empty");
6239
6240 }
6241 #endif // ASSERT
6242 if (_bit_map->isMarked(addr)) {
6243 // Obj arrays are precisely marked, non-arrays are not;
6244 // so we scan objArrays precisely and non-arrays in their
6245 // entirety.
6246 if (p->is_objArray()) {
6247 is_obj_array = true;
6248 if (_parallel) {
6249 p->oop_iterate(_par_scan_closure, mr);
6250 } else {
6251 p->oop_iterate(_scan_closure, mr);
6252 }
6253 } else {
6254 if (_parallel) {
6255 p->oop_iterate(_par_scan_closure);
6256 } else {
6257 p->oop_iterate(_scan_closure);
6258 }
6259 }
6260 }
6261 #ifdef ASSERT
6262 if (!_parallel) {
6263 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)");
6264 assert(_collector->overflow_list_is_empty(),
6265 "overflow list should be empty");
6266
6267 }
6268 #endif // ASSERT
6269 return is_obj_array;
6270 }
6271
6272 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector,
6273 MemRegion span,
6274 CMSBitMap* bitMap, CMSMarkStack* markStack,
6275 bool should_yield, bool verifying):
6276 _collector(collector),
6277 _span(span),
6278 _bitMap(bitMap),
6279 _mut(&collector->_modUnionTable),
6280 _markStack(markStack),
6281 _yield(should_yield),
6282 _skipBits(0)
6283 {
6284 assert(_markStack->isEmpty(), "stack should be empty");
6285 _finger = _bitMap->startWord();
6286 _threshold = _finger;
6287 assert(_collector->_restart_addr == NULL, "Sanity check");
6288 assert(_span.contains(_finger), "Out of bounds _finger?");
6289 DEBUG_ONLY(_verifying = verifying;)
6290 }
6291
6292 void MarkFromRootsClosure::reset(HeapWord* addr) {
6293 assert(_markStack->isEmpty(), "would cause duplicates on stack");
6294 assert(_span.contains(addr), "Out of bounds _finger?");
6295 _finger = addr;
6296 _threshold = align_up(_finger, CardTable::card_size);
6297 }
6298
6299 // Should revisit to see if this should be restructured for
6300 // greater efficiency.
6301 bool MarkFromRootsClosure::do_bit(size_t offset) {
6302 if (_skipBits > 0) {
6303 _skipBits--;
6304 return true;
6305 }
6306 // convert offset into a HeapWord*
6307 HeapWord* addr = _bitMap->startWord() + offset;
6308 assert(_bitMap->endWord() && addr < _bitMap->endWord(),
6309 "address out of range");
6310 assert(_bitMap->isMarked(addr), "tautology");
6311 if (_bitMap->isMarked(addr+1)) {
6312 // this is an allocated but not yet initialized object
6313 assert(_skipBits == 0, "tautology");
6314 _skipBits = 2; // skip next two marked bits ("Printezis-marks")
6315 oop p = oop(addr);
6316 if (p->klass_or_null_acquire() == NULL) {
6317 DEBUG_ONLY(if (!_verifying) {)
6318 // We re-dirty the cards on which this object lies and increase
6319 // the _threshold so that we'll come back to scan this object
6320 // during the preclean or remark phase. (CMSCleanOnEnter)
6321 if (CMSCleanOnEnter) {
6322 size_t sz = _collector->block_size_using_printezis_bits(addr);
6323 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
6324 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6325 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6326 // Bump _threshold to end_card_addr; note that
6327 // _threshold cannot possibly exceed end_card_addr, anyhow.
6328 // This prevents future clearing of the card as the scan proceeds
6329 // to the right.
6330 assert(_threshold <= end_card_addr,
6331 "Because we are just scanning into this object");
6332 if (_threshold < end_card_addr) {
6333 _threshold = end_card_addr;
6334 }
6335 if (p->klass_or_null_acquire() != NULL) {
6336 // Redirty the range of cards...
6337 _mut->mark_range(redirty_range);
6338 } // ...else the setting of klass will dirty the card anyway.
6339 }
6340 DEBUG_ONLY(})
6341 return true;
6342 }
6343 }
6344 scanOopsInOop(addr);
6345 return true;
6346 }
6347
6348 // We take a break if we've been at this for a while,
6349 // so as to avoid monopolizing the locks involved.
6350 void MarkFromRootsClosure::do_yield_work() {
6351 // First give up the locks, then yield, then re-lock
6352 // We should probably use a constructor/destructor idiom to
6353 // do this unlock/lock or modify the MutexUnlocker class to
6354 // serve our purpose. XXX
6355 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6356 "CMS thread should hold CMS token");
6357 assert_lock_strong(_bitMap->lock());
6358 _bitMap->lock()->unlock();
6359 ConcurrentMarkSweepThread::desynchronize(true);
6360 _collector->stopTimer();
6361 _collector->incrementYields();
6362
6363 // See the comment in coordinator_yield()
6364 for (unsigned i = 0; i < CMSYieldSleepCount &&
6365 ConcurrentMarkSweepThread::should_yield() &&
6366 !CMSCollector::foregroundGCIsActive(); ++i) {
6367 os::sleep(Thread::current(), 1, false);
6368 }
6369
6370 ConcurrentMarkSweepThread::synchronize(true);
6371 _bitMap->lock()->lock_without_safepoint_check();
6372 _collector->startTimer();
6373 }
6374
6375 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) {
6376 assert(_bitMap->isMarked(ptr), "expected bit to be set");
6377 assert(_markStack->isEmpty(),
6378 "should drain stack to limit stack usage");
6379 // convert ptr to an oop preparatory to scanning
6380 oop obj = oop(ptr);
6381 // Ignore mark word in verification below, since we
6382 // may be running concurrent with mutators.
6383 assert(oopDesc::is_oop(obj, true), "should be an oop");
6384 assert(_finger <= ptr, "_finger runneth ahead");
6385 // advance the finger to right end of this object
6386 _finger = ptr + obj->size();
6387 assert(_finger > ptr, "we just incremented it above");
6388 // On large heaps, it may take us some time to get through
6389 // the marking phase. During
6390 // this time it's possible that a lot of mutations have
6391 // accumulated in the card table and the mod union table --
6392 // these mutation records are redundant until we have
6393 // actually traced into the corresponding card.
6394 // Here, we check whether advancing the finger would make
6395 // us cross into a new card, and if so clear corresponding
6396 // cards in the MUT (preclean them in the card-table in the
6397 // future).
6398
6399 DEBUG_ONLY(if (!_verifying) {)
6400 // The clean-on-enter optimization is disabled by default,
6401 // until we fix 6178663.
6402 if (CMSCleanOnEnter && (_finger > _threshold)) {
6403 // [_threshold, _finger) represents the interval
6404 // of cards to be cleared in MUT (or precleaned in card table).
6405 // The set of cards to be cleared is all those that overlap
6406 // with the interval [_threshold, _finger); note that
6407 // _threshold is always kept card-aligned but _finger isn't
6408 // always card-aligned.
6409 HeapWord* old_threshold = _threshold;
6410 assert(is_aligned(old_threshold, CardTable::card_size),
6411 "_threshold should always be card-aligned");
6412 _threshold = align_up(_finger, CardTable::card_size);
6413 MemRegion mr(old_threshold, _threshold);
6414 assert(!mr.is_empty(), "Control point invariant");
6415 assert(_span.contains(mr), "Should clear within span");
6416 _mut->clear_range(mr);
6417 }
6418 DEBUG_ONLY(})
6419 // Note: the finger doesn't advance while we drain
6420 // the stack below.
6421 PushOrMarkClosure pushOrMarkClosure(_collector,
6422 _span, _bitMap, _markStack,
6423 _finger, this);
6424 bool res = _markStack->push(obj);
6425 assert(res, "Empty non-zero size stack should have space for single push");
6426 while (!_markStack->isEmpty()) {
6427 oop new_oop = _markStack->pop();
6428 // Skip verifying header mark word below because we are
6429 // running concurrent with mutators.
6430 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6431 // now scan this oop's oops
6432 new_oop->oop_iterate(&pushOrMarkClosure);
6433 do_yield_check();
6434 }
6435 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition");
6436 }
6437
6438 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task,
6439 CMSCollector* collector, MemRegion span,
6440 CMSBitMap* bit_map,
6441 OopTaskQueue* work_queue,
6442 CMSMarkStack* overflow_stack):
6443 _collector(collector),
6444 _whole_span(collector->_span),
6445 _span(span),
6446 _bit_map(bit_map),
6447 _mut(&collector->_modUnionTable),
6448 _work_queue(work_queue),
6449 _overflow_stack(overflow_stack),
6450 _skip_bits(0),
6451 _task(task)
6452 {
6453 assert(_work_queue->size() == 0, "work_queue should be empty");
6454 _finger = span.start();
6455 _threshold = _finger; // XXX Defer clear-on-enter optimization for now
6456 assert(_span.contains(_finger), "Out of bounds _finger?");
6457 }
6458
6459 // Should revisit to see if this should be restructured for
6460 // greater efficiency.
6461 bool ParMarkFromRootsClosure::do_bit(size_t offset) {
6462 if (_skip_bits > 0) {
6463 _skip_bits--;
6464 return true;
6465 }
6466 // convert offset into a HeapWord*
6467 HeapWord* addr = _bit_map->startWord() + offset;
6468 assert(_bit_map->endWord() && addr < _bit_map->endWord(),
6469 "address out of range");
6470 assert(_bit_map->isMarked(addr), "tautology");
6471 if (_bit_map->isMarked(addr+1)) {
6472 // this is an allocated object that might not yet be initialized
6473 assert(_skip_bits == 0, "tautology");
6474 _skip_bits = 2; // skip next two marked bits ("Printezis-marks")
6475 oop p = oop(addr);
6476 if (p->klass_or_null_acquire() == NULL) {
6477 // in the case of Clean-on-Enter optimization, redirty card
6478 // and avoid clearing card by increasing the threshold.
6479 return true;
6480 }
6481 }
6482 scan_oops_in_oop(addr);
6483 return true;
6484 }
6485
6486 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) {
6487 assert(_bit_map->isMarked(ptr), "expected bit to be set");
6488 // Should we assert that our work queue is empty or
6489 // below some drain limit?
6490 assert(_work_queue->size() == 0,
6491 "should drain stack to limit stack usage");
6492 // convert ptr to an oop preparatory to scanning
6493 oop obj = oop(ptr);
6494 // Ignore mark word in verification below, since we
6495 // may be running concurrent with mutators.
6496 assert(oopDesc::is_oop(obj, true), "should be an oop");
6497 assert(_finger <= ptr, "_finger runneth ahead");
6498 // advance the finger to right end of this object
6499 _finger = ptr + obj->size();
6500 assert(_finger > ptr, "we just incremented it above");
6501 // On large heaps, it may take us some time to get through
6502 // the marking phase. During
6503 // this time it's possible that a lot of mutations have
6504 // accumulated in the card table and the mod union table --
6505 // these mutation records are redundant until we have
6506 // actually traced into the corresponding card.
6507 // Here, we check whether advancing the finger would make
6508 // us cross into a new card, and if so clear corresponding
6509 // cards in the MUT (preclean them in the card-table in the
6510 // future).
6511
6512 // The clean-on-enter optimization is disabled by default,
6513 // until we fix 6178663.
6514 if (CMSCleanOnEnter && (_finger > _threshold)) {
6515 // [_threshold, _finger) represents the interval
6516 // of cards to be cleared in MUT (or precleaned in card table).
6517 // The set of cards to be cleared is all those that overlap
6518 // with the interval [_threshold, _finger); note that
6519 // _threshold is always kept card-aligned but _finger isn't
6520 // always card-aligned.
6521 HeapWord* old_threshold = _threshold;
6522 assert(is_aligned(old_threshold, CardTable::card_size),
6523 "_threshold should always be card-aligned");
6524 _threshold = align_up(_finger, CardTable::card_size);
6525 MemRegion mr(old_threshold, _threshold);
6526 assert(!mr.is_empty(), "Control point invariant");
6527 assert(_span.contains(mr), "Should clear within span"); // _whole_span ??
6528 _mut->clear_range(mr);
6529 }
6530
6531 // Note: the local finger doesn't advance while we drain
6532 // the stack below, but the global finger sure can and will.
6533 HeapWord* volatile* gfa = _task->global_finger_addr();
6534 ParPushOrMarkClosure pushOrMarkClosure(_collector,
6535 _span, _bit_map,
6536 _work_queue,
6537 _overflow_stack,
6538 _finger,
6539 gfa, this);
6540 bool res = _work_queue->push(obj); // overflow could occur here
6541 assert(res, "Will hold once we use workqueues");
6542 while (true) {
6543 oop new_oop;
6544 if (!_work_queue->pop_local(new_oop)) {
6545 // We emptied our work_queue; check if there's stuff that can
6546 // be gotten from the overflow stack.
6547 if (CMSConcMarkingTask::get_work_from_overflow_stack(
6548 _overflow_stack, _work_queue)) {
6549 do_yield_check();
6550 continue;
6551 } else { // done
6552 break;
6553 }
6554 }
6555 // Skip verifying header mark word below because we are
6556 // running concurrent with mutators.
6557 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop");
6558 // now scan this oop's oops
6559 new_oop->oop_iterate(&pushOrMarkClosure);
6560 do_yield_check();
6561 }
6562 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition");
6563 }
6564
6565 // Yield in response to a request from VM Thread or
6566 // from mutators.
6567 void ParMarkFromRootsClosure::do_yield_work() {
6568 assert(_task != NULL, "sanity");
6569 _task->yield();
6570 }
6571
6572 // A variant of the above used for verifying CMS marking work.
6573 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector,
6574 MemRegion span,
6575 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6576 CMSMarkStack* mark_stack):
6577 _collector(collector),
6578 _span(span),
6579 _verification_bm(verification_bm),
6580 _cms_bm(cms_bm),
6581 _mark_stack(mark_stack),
6582 _pam_verify_closure(collector, span, verification_bm, cms_bm,
6583 mark_stack)
6584 {
6585 assert(_mark_stack->isEmpty(), "stack should be empty");
6586 _finger = _verification_bm->startWord();
6587 assert(_collector->_restart_addr == NULL, "Sanity check");
6588 assert(_span.contains(_finger), "Out of bounds _finger?");
6589 }
6590
6591 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) {
6592 assert(_mark_stack->isEmpty(), "would cause duplicates on stack");
6593 assert(_span.contains(addr), "Out of bounds _finger?");
6594 _finger = addr;
6595 }
6596
6597 // Should revisit to see if this should be restructured for
6598 // greater efficiency.
6599 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) {
6600 // convert offset into a HeapWord*
6601 HeapWord* addr = _verification_bm->startWord() + offset;
6602 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(),
6603 "address out of range");
6604 assert(_verification_bm->isMarked(addr), "tautology");
6605 assert(_cms_bm->isMarked(addr), "tautology");
6606
6607 assert(_mark_stack->isEmpty(),
6608 "should drain stack to limit stack usage");
6609 // convert addr to an oop preparatory to scanning
6610 oop obj = oop(addr);
6611 assert(oopDesc::is_oop(obj), "should be an oop");
6612 assert(_finger <= addr, "_finger runneth ahead");
6613 // advance the finger to right end of this object
6614 _finger = addr + obj->size();
6615 assert(_finger > addr, "we just incremented it above");
6616 // Note: the finger doesn't advance while we drain
6617 // the stack below.
6618 bool res = _mark_stack->push(obj);
6619 assert(res, "Empty non-zero size stack should have space for single push");
6620 while (!_mark_stack->isEmpty()) {
6621 oop new_oop = _mark_stack->pop();
6622 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop");
6623 // now scan this oop's oops
6624 new_oop->oop_iterate(&_pam_verify_closure);
6625 }
6626 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition");
6627 return true;
6628 }
6629
6630 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure(
6631 CMSCollector* collector, MemRegion span,
6632 CMSBitMap* verification_bm, CMSBitMap* cms_bm,
6633 CMSMarkStack* mark_stack):
6634 MetadataAwareOopClosure(collector->ref_processor()),
6635 _collector(collector),
6636 _span(span),
6637 _verification_bm(verification_bm),
6638 _cms_bm(cms_bm),
6639 _mark_stack(mark_stack)
6640 { }
6641
6642 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) {
6643 oop obj = RawAccess<>::oop_load(p);
6644 do_oop(obj);
6645 }
6646
6647 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6648 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); }
6649
6650 // Upon stack overflow, we discard (part of) the stack,
6651 // remembering the least address amongst those discarded
6652 // in CMSCollector's _restart_address.
6653 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) {
6654 // Remember the least grey address discarded
6655 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost);
6656 _collector->lower_restart_addr(ra);
6657 _mark_stack->reset(); // discard stack contents
6658 _mark_stack->expand(); // expand the stack if possible
6659 }
6660
6661 void PushAndMarkVerifyClosure::do_oop(oop obj) {
6662 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6663 HeapWord* addr = (HeapWord*)obj;
6664 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) {
6665 // Oop lies in _span and isn't yet grey or black
6666 _verification_bm->mark(addr); // now grey
6667 if (!_cms_bm->isMarked(addr)) {
6668 Log(gc, verify) log;
6669 ResourceMark rm;
6670 LogStream ls(log.error());
6671 oop(addr)->print_on(&ls);
6672 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr));
6673 fatal("... aborting");
6674 }
6675
6676 if (!_mark_stack->push(obj)) { // stack overflow
6677 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity());
6678 assert(_mark_stack->isFull(), "Else push should have succeeded");
6679 handle_stack_overflow(addr);
6680 }
6681 // anything including and to the right of _finger
6682 // will be scanned as we iterate over the remainder of the
6683 // bit map
6684 }
6685 }
6686
6687 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector,
6688 MemRegion span,
6689 CMSBitMap* bitMap, CMSMarkStack* markStack,
6690 HeapWord* finger, MarkFromRootsClosure* parent) :
6691 MetadataAwareOopClosure(collector->ref_processor()),
6692 _collector(collector),
6693 _span(span),
6694 _bitMap(bitMap),
6695 _markStack(markStack),
6696 _finger(finger),
6697 _parent(parent)
6698 { }
6699
6700 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector,
6701 MemRegion span,
6702 CMSBitMap* bit_map,
6703 OopTaskQueue* work_queue,
6704 CMSMarkStack* overflow_stack,
6705 HeapWord* finger,
6706 HeapWord* volatile* global_finger_addr,
6707 ParMarkFromRootsClosure* parent) :
6708 MetadataAwareOopClosure(collector->ref_processor()),
6709 _collector(collector),
6710 _whole_span(collector->_span),
6711 _span(span),
6712 _bit_map(bit_map),
6713 _work_queue(work_queue),
6714 _overflow_stack(overflow_stack),
6715 _finger(finger),
6716 _global_finger_addr(global_finger_addr),
6717 _parent(parent)
6718 { }
6719
6720 // Assumes thread-safe access by callers, who are
6721 // responsible for mutual exclusion.
6722 void CMSCollector::lower_restart_addr(HeapWord* low) {
6723 assert(_span.contains(low), "Out of bounds addr");
6724 if (_restart_addr == NULL) {
6725 _restart_addr = low;
6726 } else {
6727 _restart_addr = MIN2(_restart_addr, low);
6728 }
6729 }
6730
6731 // Upon stack overflow, we discard (part of) the stack,
6732 // remembering the least address amongst those discarded
6733 // in CMSCollector's _restart_address.
6734 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6735 // Remember the least grey address discarded
6736 HeapWord* ra = (HeapWord*)_markStack->least_value(lost);
6737 _collector->lower_restart_addr(ra);
6738 _markStack->reset(); // discard stack contents
6739 _markStack->expand(); // expand the stack if possible
6740 }
6741
6742 // Upon stack overflow, we discard (part of) the stack,
6743 // remembering the least address amongst those discarded
6744 // in CMSCollector's _restart_address.
6745 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) {
6746 // We need to do this under a mutex to prevent other
6747 // workers from interfering with the work done below.
6748 MutexLockerEx ml(_overflow_stack->par_lock(),
6749 Mutex::_no_safepoint_check_flag);
6750 // Remember the least grey address discarded
6751 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost);
6752 _collector->lower_restart_addr(ra);
6753 _overflow_stack->reset(); // discard stack contents
6754 _overflow_stack->expand(); // expand the stack if possible
6755 }
6756
6757 void PushOrMarkClosure::do_oop(oop obj) {
6758 // Ignore mark word because we are running concurrent with mutators.
6759 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6760 HeapWord* addr = (HeapWord*)obj;
6761 if (_span.contains(addr) && !_bitMap->isMarked(addr)) {
6762 // Oop lies in _span and isn't yet grey or black
6763 _bitMap->mark(addr); // now grey
6764 if (addr < _finger) {
6765 // the bit map iteration has already either passed, or
6766 // sampled, this bit in the bit map; we'll need to
6767 // use the marking stack to scan this oop's oops.
6768 bool simulate_overflow = false;
6769 NOT_PRODUCT(
6770 if (CMSMarkStackOverflowALot &&
6771 _collector->simulate_overflow()) {
6772 // simulate a stack overflow
6773 simulate_overflow = true;
6774 }
6775 )
6776 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow
6777 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity());
6778 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded");
6779 handle_stack_overflow(addr);
6780 }
6781 }
6782 // anything including and to the right of _finger
6783 // will be scanned as we iterate over the remainder of the
6784 // bit map
6785 do_yield_check();
6786 }
6787 }
6788
6789 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); }
6790 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); }
6791
6792 void ParPushOrMarkClosure::do_oop(oop obj) {
6793 // Ignore mark word because we are running concurrent with mutators.
6794 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6795 HeapWord* addr = (HeapWord*)obj;
6796 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) {
6797 // Oop lies in _span and isn't yet grey or black
6798 // We read the global_finger (volatile read) strictly after marking oop
6799 bool res = _bit_map->par_mark(addr); // now grey
6800 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr;
6801 // Should we push this marked oop on our stack?
6802 // -- if someone else marked it, nothing to do
6803 // -- if target oop is above global finger nothing to do
6804 // -- if target oop is in chunk and above local finger
6805 // then nothing to do
6806 // -- else push on work queue
6807 if ( !res // someone else marked it, they will deal with it
6808 || (addr >= *gfa) // will be scanned in a later task
6809 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk
6810 return;
6811 }
6812 // the bit map iteration has already either passed, or
6813 // sampled, this bit in the bit map; we'll need to
6814 // use the marking stack to scan this oop's oops.
6815 bool simulate_overflow = false;
6816 NOT_PRODUCT(
6817 if (CMSMarkStackOverflowALot &&
6818 _collector->simulate_overflow()) {
6819 // simulate a stack overflow
6820 simulate_overflow = true;
6821 }
6822 )
6823 if (simulate_overflow ||
6824 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) {
6825 // stack overflow
6826 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity());
6827 // We cannot assert that the overflow stack is full because
6828 // it may have been emptied since.
6829 assert(simulate_overflow ||
6830 _work_queue->size() == _work_queue->max_elems(),
6831 "Else push should have succeeded");
6832 handle_stack_overflow(addr);
6833 }
6834 do_yield_check();
6835 }
6836 }
6837
6838 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6839 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); }
6840
6841 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector,
6842 MemRegion span,
6843 ReferenceDiscoverer* rd,
6844 CMSBitMap* bit_map,
6845 CMSBitMap* mod_union_table,
6846 CMSMarkStack* mark_stack,
6847 bool concurrent_precleaning):
6848 MetadataAwareOopClosure(rd),
6849 _collector(collector),
6850 _span(span),
6851 _bit_map(bit_map),
6852 _mod_union_table(mod_union_table),
6853 _mark_stack(mark_stack),
6854 _concurrent_precleaning(concurrent_precleaning)
6855 {
6856 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL");
6857 }
6858
6859 // Grey object rescan during pre-cleaning and second checkpoint phases --
6860 // the non-parallel version (the parallel version appears further below.)
6861 void PushAndMarkClosure::do_oop(oop obj) {
6862 // Ignore mark word verification. If during concurrent precleaning,
6863 // the object monitor may be locked. If during the checkpoint
6864 // phases, the object may already have been reached by a different
6865 // path and may be at the end of the global overflow list (so
6866 // the mark word may be NULL).
6867 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */),
6868 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6869 HeapWord* addr = (HeapWord*)obj;
6870 // Check if oop points into the CMS generation
6871 // and is not marked
6872 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6873 // a white object ...
6874 _bit_map->mark(addr); // ... now grey
6875 // push on the marking stack (grey set)
6876 bool simulate_overflow = false;
6877 NOT_PRODUCT(
6878 if (CMSMarkStackOverflowALot &&
6879 _collector->simulate_overflow()) {
6880 // simulate a stack overflow
6881 simulate_overflow = true;
6882 }
6883 )
6884 if (simulate_overflow || !_mark_stack->push(obj)) {
6885 if (_concurrent_precleaning) {
6886 // During precleaning we can just dirty the appropriate card(s)
6887 // in the mod union table, thus ensuring that the object remains
6888 // in the grey set and continue. In the case of object arrays
6889 // we need to dirty all of the cards that the object spans,
6890 // since the rescan of object arrays will be limited to the
6891 // dirty cards.
6892 // Note that no one can be interfering with us in this action
6893 // of dirtying the mod union table, so no locking or atomics
6894 // are required.
6895 if (obj->is_objArray()) {
6896 size_t sz = obj->size();
6897 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
6898 MemRegion redirty_range = MemRegion(addr, end_card_addr);
6899 assert(!redirty_range.is_empty(), "Arithmetical tautology");
6900 _mod_union_table->mark_range(redirty_range);
6901 } else {
6902 _mod_union_table->mark(addr);
6903 }
6904 _collector->_ser_pmc_preclean_ovflw++;
6905 } else {
6906 // During the remark phase, we need to remember this oop
6907 // in the overflow list.
6908 _collector->push_on_overflow_list(obj);
6909 _collector->_ser_pmc_remark_ovflw++;
6910 }
6911 }
6912 }
6913 }
6914
6915 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector,
6916 MemRegion span,
6917 ReferenceDiscoverer* rd,
6918 CMSBitMap* bit_map,
6919 OopTaskQueue* work_queue):
6920 MetadataAwareOopClosure(rd),
6921 _collector(collector),
6922 _span(span),
6923 _bit_map(bit_map),
6924 _work_queue(work_queue)
6925 {
6926 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL");
6927 }
6928
6929 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); }
6930 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); }
6931
6932 // Grey object rescan during second checkpoint phase --
6933 // the parallel version.
6934 void ParPushAndMarkClosure::do_oop(oop obj) {
6935 // In the assert below, we ignore the mark word because
6936 // this oop may point to an already visited object that is
6937 // on the overflow stack (in which case the mark word has
6938 // been hijacked for chaining into the overflow stack --
6939 // if this is the last object in the overflow stack then
6940 // its mark word will be NULL). Because this object may
6941 // have been subsequently popped off the global overflow
6942 // stack, and the mark word possibly restored to the prototypical
6943 // value, by the time we get to examined this failing assert in
6944 // the debugger, is_oop_or_null(false) may subsequently start
6945 // to hold.
6946 assert(oopDesc::is_oop_or_null(obj, true),
6947 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj));
6948 HeapWord* addr = (HeapWord*)obj;
6949 // Check if oop points into the CMS generation
6950 // and is not marked
6951 if (_span.contains(addr) && !_bit_map->isMarked(addr)) {
6952 // a white object ...
6953 // If we manage to "claim" the object, by being the
6954 // first thread to mark it, then we push it on our
6955 // marking stack
6956 if (_bit_map->par_mark(addr)) { // ... now grey
6957 // push on work queue (grey set)
6958 bool simulate_overflow = false;
6959 NOT_PRODUCT(
6960 if (CMSMarkStackOverflowALot &&
6961 _collector->par_simulate_overflow()) {
6962 // simulate a stack overflow
6963 simulate_overflow = true;
6964 }
6965 )
6966 if (simulate_overflow || !_work_queue->push(obj)) {
6967 _collector->par_push_on_overflow_list(obj);
6968 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS
6969 }
6970 } // Else, some other thread got there first
6971 }
6972 }
6973
6974 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6975 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); }
6976
6977 void CMSPrecleanRefsYieldClosure::do_yield_work() {
6978 Mutex* bml = _collector->bitMapLock();
6979 assert_lock_strong(bml);
6980 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
6981 "CMS thread should hold CMS token");
6982
6983 bml->unlock();
6984 ConcurrentMarkSweepThread::desynchronize(true);
6985
6986 _collector->stopTimer();
6987 _collector->incrementYields();
6988
6989 // See the comment in coordinator_yield()
6990 for (unsigned i = 0; i < CMSYieldSleepCount &&
6991 ConcurrentMarkSweepThread::should_yield() &&
6992 !CMSCollector::foregroundGCIsActive(); ++i) {
6993 os::sleep(Thread::current(), 1, false);
6994 }
6995
6996 ConcurrentMarkSweepThread::synchronize(true);
6997 bml->lock();
6998
6999 _collector->startTimer();
7000 }
7001
7002 bool CMSPrecleanRefsYieldClosure::should_return() {
7003 if (ConcurrentMarkSweepThread::should_yield()) {
7004 do_yield_work();
7005 }
7006 return _collector->foregroundGCIsActive();
7007 }
7008
7009 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) {
7010 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0,
7011 "mr should be aligned to start at a card boundary");
7012 // We'd like to assert:
7013 // assert(mr.word_size()%CardTable::card_size_in_words == 0,
7014 // "mr should be a range of cards");
7015 // However, that would be too strong in one case -- the last
7016 // partition ends at _unallocated_block which, in general, can be
7017 // an arbitrary boundary, not necessarily card aligned.
7018 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words;
7019 _space->object_iterate_mem(mr, &_scan_cl);
7020 }
7021
7022 SweepClosure::SweepClosure(CMSCollector* collector,
7023 ConcurrentMarkSweepGeneration* g,
7024 CMSBitMap* bitMap, bool should_yield) :
7025 _collector(collector),
7026 _g(g),
7027 _sp(g->cmsSpace()),
7028 _limit(_sp->sweep_limit()),
7029 _freelistLock(_sp->freelistLock()),
7030 _bitMap(bitMap),
7031 _yield(should_yield),
7032 _inFreeRange(false), // No free range at beginning of sweep
7033 _freeRangeInFreeLists(false), // No free range at beginning of sweep
7034 _lastFreeRangeCoalesced(false),
7035 _freeFinger(g->used_region().start())
7036 {
7037 NOT_PRODUCT(
7038 _numObjectsFreed = 0;
7039 _numWordsFreed = 0;
7040 _numObjectsLive = 0;
7041 _numWordsLive = 0;
7042 _numObjectsAlreadyFree = 0;
7043 _numWordsAlreadyFree = 0;
7044 _last_fc = NULL;
7045
7046 _sp->initializeIndexedFreeListArrayReturnedBytes();
7047 _sp->dictionary()->initialize_dict_returned_bytes();
7048 )
7049 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7050 "sweep _limit out of bounds");
7051 log_develop_trace(gc, sweep)("====================");
7052 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit));
7053 }
7054
7055 void SweepClosure::print_on(outputStream* st) const {
7056 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")",
7057 p2i(_sp->bottom()), p2i(_sp->end()));
7058 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit));
7059 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger));
7060 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));)
7061 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d",
7062 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced);
7063 }
7064
7065 #ifndef PRODUCT
7066 // Assertion checking only: no useful work in product mode --
7067 // however, if any of the flags below become product flags,
7068 // you may need to review this code to see if it needs to be
7069 // enabled in product mode.
7070 SweepClosure::~SweepClosure() {
7071 assert_lock_strong(_freelistLock);
7072 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7073 "sweep _limit out of bounds");
7074 if (inFreeRange()) {
7075 Log(gc, sweep) log;
7076 log.error("inFreeRange() should have been reset; dumping state of SweepClosure");
7077 ResourceMark rm;
7078 LogStream ls(log.error());
7079 print_on(&ls);
7080 ShouldNotReachHere();
7081 }
7082
7083 if (log_is_enabled(Debug, gc, sweep)) {
7084 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7085 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord));
7086 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes",
7087 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord));
7088 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord);
7089 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes);
7090 }
7091
7092 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) {
7093 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes();
7094 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes();
7095 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes;
7096 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes",
7097 returned_bytes, indexListReturnedBytes, dict_returned_bytes);
7098 }
7099 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit));
7100 log_develop_trace(gc, sweep)("================");
7101 }
7102 #endif // PRODUCT
7103
7104 void SweepClosure::initialize_free_range(HeapWord* freeFinger,
7105 bool freeRangeInFreeLists) {
7106 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)",
7107 p2i(freeFinger), freeRangeInFreeLists);
7108 assert(!inFreeRange(), "Trampling existing free range");
7109 set_inFreeRange(true);
7110 set_lastFreeRangeCoalesced(false);
7111
7112 set_freeFinger(freeFinger);
7113 set_freeRangeInFreeLists(freeRangeInFreeLists);
7114 if (CMSTestInFreeList) {
7115 if (freeRangeInFreeLists) {
7116 FreeChunk* fc = (FreeChunk*) freeFinger;
7117 assert(fc->is_free(), "A chunk on the free list should be free.");
7118 assert(fc->size() > 0, "Free range should have a size");
7119 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists");
7120 }
7121 }
7122 }
7123
7124 // Note that the sweeper runs concurrently with mutators. Thus,
7125 // it is possible for direct allocation in this generation to happen
7126 // in the middle of the sweep. Note that the sweeper also coalesces
7127 // contiguous free blocks. Thus, unless the sweeper and the allocator
7128 // synchronize appropriately freshly allocated blocks may get swept up.
7129 // This is accomplished by the sweeper locking the free lists while
7130 // it is sweeping. Thus blocks that are determined to be free are
7131 // indeed free. There is however one additional complication:
7132 // blocks that have been allocated since the final checkpoint and
7133 // mark, will not have been marked and so would be treated as
7134 // unreachable and swept up. To prevent this, the allocator marks
7135 // the bit map when allocating during the sweep phase. This leads,
7136 // however, to a further complication -- objects may have been allocated
7137 // but not yet initialized -- in the sense that the header isn't yet
7138 // installed. The sweeper can not then determine the size of the block
7139 // in order to skip over it. To deal with this case, we use a technique
7140 // (due to Printezis) to encode such uninitialized block sizes in the
7141 // bit map. Since the bit map uses a bit per every HeapWord, but the
7142 // CMS generation has a minimum object size of 3 HeapWords, it follows
7143 // that "normal marks" won't be adjacent in the bit map (there will
7144 // always be at least two 0 bits between successive 1 bits). We make use
7145 // of these "unused" bits to represent uninitialized blocks -- the bit
7146 // corresponding to the start of the uninitialized object and the next
7147 // bit are both set. Finally, a 1 bit marks the end of the object that
7148 // started with the two consecutive 1 bits to indicate its potentially
7149 // uninitialized state.
7150
7151 size_t SweepClosure::do_blk_careful(HeapWord* addr) {
7152 FreeChunk* fc = (FreeChunk*)addr;
7153 size_t res;
7154
7155 // Check if we are done sweeping. Below we check "addr >= _limit" rather
7156 // than "addr == _limit" because although _limit was a block boundary when
7157 // we started the sweep, it may no longer be one because heap expansion
7158 // may have caused us to coalesce the block ending at the address _limit
7159 // with a newly expanded chunk (this happens when _limit was set to the
7160 // previous _end of the space), so we may have stepped past _limit:
7161 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740.
7162 if (addr >= _limit) { // we have swept up to or past the limit: finish up
7163 assert(_limit >= _sp->bottom() && _limit <= _sp->end(),
7164 "sweep _limit out of bounds");
7165 assert(addr < _sp->end(), "addr out of bounds");
7166 // Flush any free range we might be holding as a single
7167 // coalesced chunk to the appropriate free list.
7168 if (inFreeRange()) {
7169 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit,
7170 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger()));
7171 flush_cur_free_chunk(freeFinger(),
7172 pointer_delta(addr, freeFinger()));
7173 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]",
7174 p2i(freeFinger()), pointer_delta(addr, freeFinger()),
7175 lastFreeRangeCoalesced() ? 1 : 0);
7176 }
7177
7178 // help the iterator loop finish
7179 return pointer_delta(_sp->end(), addr);
7180 }
7181
7182 assert(addr < _limit, "sweep invariant");
7183 // check if we should yield
7184 do_yield_check(addr);
7185 if (fc->is_free()) {
7186 // Chunk that is already free
7187 res = fc->size();
7188 do_already_free_chunk(fc);
7189 debug_only(_sp->verifyFreeLists());
7190 // If we flush the chunk at hand in lookahead_and_flush()
7191 // and it's coalesced with a preceding chunk, then the
7192 // process of "mangling" the payload of the coalesced block
7193 // will cause erasure of the size information from the
7194 // (erstwhile) header of all the coalesced blocks but the
7195 // first, so the first disjunct in the assert will not hold
7196 // in that specific case (in which case the second disjunct
7197 // will hold).
7198 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit,
7199 "Otherwise the size info doesn't change at this step");
7200 NOT_PRODUCT(
7201 _numObjectsAlreadyFree++;
7202 _numWordsAlreadyFree += res;
7203 )
7204 NOT_PRODUCT(_last_fc = fc;)
7205 } else if (!_bitMap->isMarked(addr)) {
7206 // Chunk is fresh garbage
7207 res = do_garbage_chunk(fc);
7208 debug_only(_sp->verifyFreeLists());
7209 NOT_PRODUCT(
7210 _numObjectsFreed++;
7211 _numWordsFreed += res;
7212 )
7213 } else {
7214 // Chunk that is alive.
7215 res = do_live_chunk(fc);
7216 debug_only(_sp->verifyFreeLists());
7217 NOT_PRODUCT(
7218 _numObjectsLive++;
7219 _numWordsLive += res;
7220 )
7221 }
7222 return res;
7223 }
7224
7225 // For the smart allocation, record following
7226 // split deaths - a free chunk is removed from its free list because
7227 // it is being split into two or more chunks.
7228 // split birth - a free chunk is being added to its free list because
7229 // a larger free chunk has been split and resulted in this free chunk.
7230 // coal death - a free chunk is being removed from its free list because
7231 // it is being coalesced into a large free chunk.
7232 // coal birth - a free chunk is being added to its free list because
7233 // it was created when two or more free chunks where coalesced into
7234 // this free chunk.
7235 //
7236 // These statistics are used to determine the desired number of free
7237 // chunks of a given size. The desired number is chosen to be relative
7238 // to the end of a CMS sweep. The desired number at the end of a sweep
7239 // is the
7240 // count-at-end-of-previous-sweep (an amount that was enough)
7241 // - count-at-beginning-of-current-sweep (the excess)
7242 // + split-births (gains in this size during interval)
7243 // - split-deaths (demands on this size during interval)
7244 // where the interval is from the end of one sweep to the end of the
7245 // next.
7246 //
7247 // When sweeping the sweeper maintains an accumulated chunk which is
7248 // the chunk that is made up of chunks that have been coalesced. That
7249 // will be termed the left-hand chunk. A new chunk of garbage that
7250 // is being considered for coalescing will be referred to as the
7251 // right-hand chunk.
7252 //
7253 // When making a decision on whether to coalesce a right-hand chunk with
7254 // the current left-hand chunk, the current count vs. the desired count
7255 // of the left-hand chunk is considered. Also if the right-hand chunk
7256 // is near the large chunk at the end of the heap (see
7257 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the
7258 // left-hand chunk is coalesced.
7259 //
7260 // When making a decision about whether to split a chunk, the desired count
7261 // vs. the current count of the candidate to be split is also considered.
7262 // If the candidate is underpopulated (currently fewer chunks than desired)
7263 // a chunk of an overpopulated (currently more chunks than desired) size may
7264 // be chosen. The "hint" associated with a free list, if non-null, points
7265 // to a free list which may be overpopulated.
7266 //
7267
7268 void SweepClosure::do_already_free_chunk(FreeChunk* fc) {
7269 const size_t size = fc->size();
7270 // Chunks that cannot be coalesced are not in the
7271 // free lists.
7272 if (CMSTestInFreeList && !fc->cantCoalesce()) {
7273 assert(_sp->verify_chunk_in_free_list(fc),
7274 "free chunk should be in free lists");
7275 }
7276 // a chunk that is already free, should not have been
7277 // marked in the bit map
7278 HeapWord* const addr = (HeapWord*) fc;
7279 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked");
7280 // Verify that the bit map has no bits marked between
7281 // addr and purported end of this block.
7282 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7283
7284 // Some chunks cannot be coalesced under any circumstances.
7285 // See the definition of cantCoalesce().
7286 if (!fc->cantCoalesce()) {
7287 // This chunk can potentially be coalesced.
7288 // All the work is done in
7289 do_post_free_or_garbage_chunk(fc, size);
7290 // Note that if the chunk is not coalescable (the else arm
7291 // below), we unconditionally flush, without needing to do
7292 // a "lookahead," as we do below.
7293 if (inFreeRange()) lookahead_and_flush(fc, size);
7294 } else {
7295 // Code path common to both original and adaptive free lists.
7296
7297 // cant coalesce with previous block; this should be treated
7298 // as the end of a free run if any
7299 if (inFreeRange()) {
7300 // we kicked some butt; time to pick up the garbage
7301 assert(freeFinger() < addr, "freeFinger points too high");
7302 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7303 }
7304 // else, nothing to do, just continue
7305 }
7306 }
7307
7308 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) {
7309 // This is a chunk of garbage. It is not in any free list.
7310 // Add it to a free list or let it possibly be coalesced into
7311 // a larger chunk.
7312 HeapWord* const addr = (HeapWord*) fc;
7313 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7314
7315 // Verify that the bit map has no bits marked between
7316 // addr and purported end of just dead object.
7317 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size);
7318 do_post_free_or_garbage_chunk(fc, size);
7319
7320 assert(_limit >= addr + size,
7321 "A freshly garbage chunk can't possibly straddle over _limit");
7322 if (inFreeRange()) lookahead_and_flush(fc, size);
7323 return size;
7324 }
7325
7326 size_t SweepClosure::do_live_chunk(FreeChunk* fc) {
7327 HeapWord* addr = (HeapWord*) fc;
7328 // The sweeper has just found a live object. Return any accumulated
7329 // left hand chunk to the free lists.
7330 if (inFreeRange()) {
7331 assert(freeFinger() < addr, "freeFinger points too high");
7332 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7333 }
7334
7335 // This object is live: we'd normally expect this to be
7336 // an oop, and like to assert the following:
7337 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop");
7338 // However, as we commented above, this may be an object whose
7339 // header hasn't yet been initialized.
7340 size_t size;
7341 assert(_bitMap->isMarked(addr), "Tautology for this control point");
7342 if (_bitMap->isMarked(addr + 1)) {
7343 // Determine the size from the bit map, rather than trying to
7344 // compute it from the object header.
7345 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2);
7346 size = pointer_delta(nextOneAddr + 1, addr);
7347 assert(size == CompactibleFreeListSpace::adjustObjectSize(size),
7348 "alignment problem");
7349
7350 #ifdef ASSERT
7351 if (oop(addr)->klass_or_null_acquire() != NULL) {
7352 // Ignore mark word because we are running concurrent with mutators
7353 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7354 assert(size ==
7355 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()),
7356 "P-mark and computed size do not agree");
7357 }
7358 #endif
7359
7360 } else {
7361 // This should be an initialized object that's alive.
7362 assert(oop(addr)->klass_or_null_acquire() != NULL,
7363 "Should be an initialized object");
7364 // Ignore mark word because we are running concurrent with mutators
7365 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop");
7366 // Verify that the bit map has no bits marked between
7367 // addr and purported end of this block.
7368 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size());
7369 assert(size >= 3, "Necessary for Printezis marks to work");
7370 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point");
7371 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);)
7372 }
7373 return size;
7374 }
7375
7376 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc,
7377 size_t chunkSize) {
7378 // do_post_free_or_garbage_chunk() should only be called in the case
7379 // of the adaptive free list allocator.
7380 const bool fcInFreeLists = fc->is_free();
7381 assert((HeapWord*)fc <= _limit, "sweep invariant");
7382 if (CMSTestInFreeList && fcInFreeLists) {
7383 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists");
7384 }
7385
7386 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize);
7387
7388 HeapWord* const fc_addr = (HeapWord*) fc;
7389
7390 bool coalesce = false;
7391 const size_t left = pointer_delta(fc_addr, freeFinger());
7392 const size_t right = chunkSize;
7393 switch (FLSCoalescePolicy) {
7394 // numeric value forms a coalition aggressiveness metric
7395 case 0: { // never coalesce
7396 coalesce = false;
7397 break;
7398 }
7399 case 1: { // coalesce if left & right chunks on overpopulated lists
7400 coalesce = _sp->coalOverPopulated(left) &&
7401 _sp->coalOverPopulated(right);
7402 break;
7403 }
7404 case 2: { // coalesce if left chunk on overpopulated list (default)
7405 coalesce = _sp->coalOverPopulated(left);
7406 break;
7407 }
7408 case 3: { // coalesce if left OR right chunk on overpopulated list
7409 coalesce = _sp->coalOverPopulated(left) ||
7410 _sp->coalOverPopulated(right);
7411 break;
7412 }
7413 case 4: { // always coalesce
7414 coalesce = true;
7415 break;
7416 }
7417 default:
7418 ShouldNotReachHere();
7419 }
7420
7421 // Should the current free range be coalesced?
7422 // If the chunk is in a free range and either we decided to coalesce above
7423 // or the chunk is near the large block at the end of the heap
7424 // (isNearLargestChunk() returns true), then coalesce this chunk.
7425 const bool doCoalesce = inFreeRange()
7426 && (coalesce || _g->isNearLargestChunk(fc_addr));
7427 if (doCoalesce) {
7428 // Coalesce the current free range on the left with the new
7429 // chunk on the right. If either is on a free list,
7430 // it must be removed from the list and stashed in the closure.
7431 if (freeRangeInFreeLists()) {
7432 FreeChunk* const ffc = (FreeChunk*)freeFinger();
7433 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()),
7434 "Size of free range is inconsistent with chunk size.");
7435 if (CMSTestInFreeList) {
7436 assert(_sp->verify_chunk_in_free_list(ffc),
7437 "Chunk is not in free lists");
7438 }
7439 _sp->coalDeath(ffc->size());
7440 _sp->removeFreeChunkFromFreeLists(ffc);
7441 set_freeRangeInFreeLists(false);
7442 }
7443 if (fcInFreeLists) {
7444 _sp->coalDeath(chunkSize);
7445 assert(fc->size() == chunkSize,
7446 "The chunk has the wrong size or is not in the free lists");
7447 _sp->removeFreeChunkFromFreeLists(fc);
7448 }
7449 set_lastFreeRangeCoalesced(true);
7450 print_free_block_coalesced(fc);
7451 } else { // not in a free range and/or should not coalesce
7452 // Return the current free range and start a new one.
7453 if (inFreeRange()) {
7454 // In a free range but cannot coalesce with the right hand chunk.
7455 // Put the current free range into the free lists.
7456 flush_cur_free_chunk(freeFinger(),
7457 pointer_delta(fc_addr, freeFinger()));
7458 }
7459 // Set up for new free range. Pass along whether the right hand
7460 // chunk is in the free lists.
7461 initialize_free_range((HeapWord*)fc, fcInFreeLists);
7462 }
7463 }
7464
7465 // Lookahead flush:
7466 // If we are tracking a free range, and this is the last chunk that
7467 // we'll look at because its end crosses past _limit, we'll preemptively
7468 // flush it along with any free range we may be holding on to. Note that
7469 // this can be the case only for an already free or freshly garbage
7470 // chunk. If this block is an object, it can never straddle
7471 // over _limit. The "straddling" occurs when _limit is set at
7472 // the previous end of the space when this cycle started, and
7473 // a subsequent heap expansion caused the previously co-terminal
7474 // free block to be coalesced with the newly expanded portion,
7475 // thus rendering _limit a non-block-boundary making it dangerous
7476 // for the sweeper to step over and examine.
7477 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) {
7478 assert(inFreeRange(), "Should only be called if currently in a free range.");
7479 HeapWord* const eob = ((HeapWord*)fc) + chunk_size;
7480 assert(_sp->used_region().contains(eob - 1),
7481 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT
7482 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")"
7483 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")",
7484 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size);
7485 if (eob >= _limit) {
7486 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit");
7487 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block "
7488 "[" PTR_FORMAT "," PTR_FORMAT ") in space "
7489 "[" PTR_FORMAT "," PTR_FORMAT ")",
7490 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end()));
7491 // Return the storage we are tracking back into the free lists.
7492 log_develop_trace(gc, sweep)("Flushing ... ");
7493 assert(freeFinger() < eob, "Error");
7494 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger()));
7495 }
7496 }
7497
7498 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) {
7499 assert(inFreeRange(), "Should only be called if currently in a free range.");
7500 assert(size > 0,
7501 "A zero sized chunk cannot be added to the free lists.");
7502 if (!freeRangeInFreeLists()) {
7503 if (CMSTestInFreeList) {
7504 FreeChunk* fc = (FreeChunk*) chunk;
7505 fc->set_size(size);
7506 assert(!_sp->verify_chunk_in_free_list(fc),
7507 "chunk should not be in free lists yet");
7508 }
7509 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size);
7510 // A new free range is going to be starting. The current
7511 // free range has not been added to the free lists yet or
7512 // was removed so add it back.
7513 // If the current free range was coalesced, then the death
7514 // of the free range was recorded. Record a birth now.
7515 if (lastFreeRangeCoalesced()) {
7516 _sp->coalBirth(size);
7517 }
7518 _sp->addChunkAndRepairOffsetTable(chunk, size,
7519 lastFreeRangeCoalesced());
7520 } else {
7521 log_develop_trace(gc, sweep)("Already in free list: nothing to flush");
7522 }
7523 set_inFreeRange(false);
7524 set_freeRangeInFreeLists(false);
7525 }
7526
7527 // We take a break if we've been at this for a while,
7528 // so as to avoid monopolizing the locks involved.
7529 void SweepClosure::do_yield_work(HeapWord* addr) {
7530 // Return current free chunk being used for coalescing (if any)
7531 // to the appropriate freelist. After yielding, the next
7532 // free block encountered will start a coalescing range of
7533 // free blocks. If the next free block is adjacent to the
7534 // chunk just flushed, they will need to wait for the next
7535 // sweep to be coalesced.
7536 if (inFreeRange()) {
7537 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger()));
7538 }
7539
7540 // First give up the locks, then yield, then re-lock.
7541 // We should probably use a constructor/destructor idiom to
7542 // do this unlock/lock or modify the MutexUnlocker class to
7543 // serve our purpose. XXX
7544 assert_lock_strong(_bitMap->lock());
7545 assert_lock_strong(_freelistLock);
7546 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(),
7547 "CMS thread should hold CMS token");
7548 _bitMap->lock()->unlock();
7549 _freelistLock->unlock();
7550 ConcurrentMarkSweepThread::desynchronize(true);
7551 _collector->stopTimer();
7552 _collector->incrementYields();
7553
7554 // See the comment in coordinator_yield()
7555 for (unsigned i = 0; i < CMSYieldSleepCount &&
7556 ConcurrentMarkSweepThread::should_yield() &&
7557 !CMSCollector::foregroundGCIsActive(); ++i) {
7558 os::sleep(Thread::current(), 1, false);
7559 }
7560
7561 ConcurrentMarkSweepThread::synchronize(true);
7562 _freelistLock->lock();
7563 _bitMap->lock()->lock_without_safepoint_check();
7564 _collector->startTimer();
7565 }
7566
7567 #ifndef PRODUCT
7568 // This is actually very useful in a product build if it can
7569 // be called from the debugger. Compile it into the product
7570 // as needed.
7571 bool debug_verify_chunk_in_free_list(FreeChunk* fc) {
7572 return debug_cms_space->verify_chunk_in_free_list(fc);
7573 }
7574 #endif
7575
7576 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const {
7577 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")",
7578 p2i(fc), fc->size());
7579 }
7580
7581 // CMSIsAliveClosure
7582 bool CMSIsAliveClosure::do_object_b(oop obj) {
7583 HeapWord* addr = (HeapWord*)obj;
7584 return addr != NULL &&
7585 (!_span.contains(addr) || _bit_map->isMarked(addr));
7586 }
7587
7588
7589 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector,
7590 MemRegion span,
7591 CMSBitMap* bit_map, CMSMarkStack* mark_stack,
7592 bool cpc):
7593 _collector(collector),
7594 _span(span),
7595 _bit_map(bit_map),
7596 _mark_stack(mark_stack),
7597 _concurrent_precleaning(cpc) {
7598 assert(!_span.is_empty(), "Empty span could spell trouble");
7599 }
7600
7601
7602 // CMSKeepAliveClosure: the serial version
7603 void CMSKeepAliveClosure::do_oop(oop obj) {
7604 HeapWord* addr = (HeapWord*)obj;
7605 if (_span.contains(addr) &&
7606 !_bit_map->isMarked(addr)) {
7607 _bit_map->mark(addr);
7608 bool simulate_overflow = false;
7609 NOT_PRODUCT(
7610 if (CMSMarkStackOverflowALot &&
7611 _collector->simulate_overflow()) {
7612 // simulate a stack overflow
7613 simulate_overflow = true;
7614 }
7615 )
7616 if (simulate_overflow || !_mark_stack->push(obj)) {
7617 if (_concurrent_precleaning) {
7618 // We dirty the overflown object and let the remark
7619 // phase deal with it.
7620 assert(_collector->overflow_list_is_empty(), "Error");
7621 // In the case of object arrays, we need to dirty all of
7622 // the cards that the object spans. No locking or atomics
7623 // are needed since no one else can be mutating the mod union
7624 // table.
7625 if (obj->is_objArray()) {
7626 size_t sz = obj->size();
7627 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size);
7628 MemRegion redirty_range = MemRegion(addr, end_card_addr);
7629 assert(!redirty_range.is_empty(), "Arithmetical tautology");
7630 _collector->_modUnionTable.mark_range(redirty_range);
7631 } else {
7632 _collector->_modUnionTable.mark(addr);
7633 }
7634 _collector->_ser_kac_preclean_ovflw++;
7635 } else {
7636 _collector->push_on_overflow_list(obj);
7637 _collector->_ser_kac_ovflw++;
7638 }
7639 }
7640 }
7641 }
7642
7643 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7644 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); }
7645
7646 // CMSParKeepAliveClosure: a parallel version of the above.
7647 // The work queues are private to each closure (thread),
7648 // but (may be) available for stealing by other threads.
7649 void CMSParKeepAliveClosure::do_oop(oop obj) {
7650 HeapWord* addr = (HeapWord*)obj;
7651 if (_span.contains(addr) &&
7652 !_bit_map->isMarked(addr)) {
7653 // In general, during recursive tracing, several threads
7654 // may be concurrently getting here; the first one to
7655 // "tag" it, claims it.
7656 if (_bit_map->par_mark(addr)) {
7657 bool res = _work_queue->push(obj);
7658 assert(res, "Low water mark should be much less than capacity");
7659 // Do a recursive trim in the hope that this will keep
7660 // stack usage lower, but leave some oops for potential stealers
7661 trim_queue(_low_water_mark);
7662 } // Else, another thread got there first
7663 }
7664 }
7665
7666 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7667 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); }
7668
7669 void CMSParKeepAliveClosure::trim_queue(uint max) {
7670 while (_work_queue->size() > max) {
7671 oop new_oop;
7672 if (_work_queue->pop_local(new_oop)) {
7673 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop");
7674 assert(_bit_map->isMarked((HeapWord*)new_oop),
7675 "no white objects on this stack!");
7676 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7677 // iterate over the oops in this oop, marking and pushing
7678 // the ones in CMS heap (i.e. in _span).
7679 new_oop->oop_iterate(&_mark_and_push);
7680 }
7681 }
7682 }
7683
7684 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure(
7685 CMSCollector* collector,
7686 MemRegion span, CMSBitMap* bit_map,
7687 OopTaskQueue* work_queue):
7688 _collector(collector),
7689 _span(span),
7690 _bit_map(bit_map),
7691 _work_queue(work_queue) { }
7692
7693 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) {
7694 HeapWord* addr = (HeapWord*)obj;
7695 if (_span.contains(addr) &&
7696 !_bit_map->isMarked(addr)) {
7697 if (_bit_map->par_mark(addr)) {
7698 bool simulate_overflow = false;
7699 NOT_PRODUCT(
7700 if (CMSMarkStackOverflowALot &&
7701 _collector->par_simulate_overflow()) {
7702 // simulate a stack overflow
7703 simulate_overflow = true;
7704 }
7705 )
7706 if (simulate_overflow || !_work_queue->push(obj)) {
7707 _collector->par_push_on_overflow_list(obj);
7708 _collector->_par_kac_ovflw++;
7709 }
7710 } // Else another thread got there already
7711 }
7712 }
7713
7714 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7715 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); }
7716
7717 //////////////////////////////////////////////////////////////////
7718 // CMSExpansionCause /////////////////////////////
7719 //////////////////////////////////////////////////////////////////
7720 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) {
7721 switch (cause) {
7722 case _no_expansion:
7723 return "No expansion";
7724 case _satisfy_free_ratio:
7725 return "Free ratio";
7726 case _satisfy_promotion:
7727 return "Satisfy promotion";
7728 case _satisfy_allocation:
7729 return "allocation";
7730 case _allocate_par_lab:
7731 return "Par LAB";
7732 case _allocate_par_spooling_space:
7733 return "Par Spooling Space";
7734 case _adaptive_size_policy:
7735 return "Ergonomics";
7736 default:
7737 return "unknown";
7738 }
7739 }
7740
7741 void CMSDrainMarkingStackClosure::do_void() {
7742 // the max number to take from overflow list at a time
7743 const size_t num = _mark_stack->capacity()/4;
7744 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(),
7745 "Overflow list should be NULL during concurrent phases");
7746 while (!_mark_stack->isEmpty() ||
7747 // if stack is empty, check the overflow list
7748 _collector->take_from_overflow_list(num, _mark_stack)) {
7749 oop obj = _mark_stack->pop();
7750 HeapWord* addr = (HeapWord*)obj;
7751 assert(_span.contains(addr), "Should be within span");
7752 assert(_bit_map->isMarked(addr), "Should be marked");
7753 assert(oopDesc::is_oop(obj), "Should be an oop");
7754 obj->oop_iterate(_keep_alive);
7755 }
7756 }
7757
7758 void CMSParDrainMarkingStackClosure::do_void() {
7759 // drain queue
7760 trim_queue(0);
7761 }
7762
7763 // Trim our work_queue so its length is below max at return
7764 void CMSParDrainMarkingStackClosure::trim_queue(uint max) {
7765 while (_work_queue->size() > max) {
7766 oop new_oop;
7767 if (_work_queue->pop_local(new_oop)) {
7768 assert(oopDesc::is_oop(new_oop), "Expected an oop");
7769 assert(_bit_map->isMarked((HeapWord*)new_oop),
7770 "no white objects on this stack!");
7771 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop");
7772 // iterate over the oops in this oop, marking and pushing
7773 // the ones in CMS heap (i.e. in _span).
7774 new_oop->oop_iterate(&_mark_and_push);
7775 }
7776 }
7777 }
7778
7779 ////////////////////////////////////////////////////////////////////
7780 // Support for Marking Stack Overflow list handling and related code
7781 ////////////////////////////////////////////////////////////////////
7782 // Much of the following code is similar in shape and spirit to the
7783 // code used in ParNewGC. We should try and share that code
7784 // as much as possible in the future.
7785
7786 #ifndef PRODUCT
7787 // Debugging support for CMSStackOverflowALot
7788
7789 // It's OK to call this multi-threaded; the worst thing
7790 // that can happen is that we'll get a bunch of closely
7791 // spaced simulated overflows, but that's OK, in fact
7792 // probably good as it would exercise the overflow code
7793 // under contention.
7794 bool CMSCollector::simulate_overflow() {
7795 if (_overflow_counter-- <= 0) { // just being defensive
7796 _overflow_counter = CMSMarkStackOverflowInterval;
7797 return true;
7798 } else {
7799 return false;
7800 }
7801 }
7802
7803 bool CMSCollector::par_simulate_overflow() {
7804 return simulate_overflow();
7805 }
7806 #endif
7807
7808 // Single-threaded
7809 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
7810 assert(stack->isEmpty(), "Expected precondition");
7811 assert(stack->capacity() > num, "Shouldn't bite more than can chew");
7812 size_t i = num;
7813 oop cur = _overflow_list;
7814 const markOop proto = markOopDesc::prototype();
7815 NOT_PRODUCT(ssize_t n = 0;)
7816 for (oop next; i > 0 && cur != NULL; cur = next, i--) {
7817 next = oop(cur->mark());
7818 cur->set_mark(proto); // until proven otherwise
7819 assert(oopDesc::is_oop(cur), "Should be an oop");
7820 bool res = stack->push(cur);
7821 assert(res, "Bit off more than can chew?");
7822 NOT_PRODUCT(n++;)
7823 }
7824 _overflow_list = cur;
7825 #ifndef PRODUCT
7826 assert(_num_par_pushes >= n, "Too many pops?");
7827 _num_par_pushes -=n;
7828 #endif
7829 return !stack->isEmpty();
7830 }
7831
7832 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
7833 // (MT-safe) Get a prefix of at most "num" from the list.
7834 // The overflow list is chained through the mark word of
7835 // each object in the list. We fetch the entire list,
7836 // break off a prefix of the right size and return the
7837 // remainder. If other threads try to take objects from
7838 // the overflow list at that time, they will wait for
7839 // some time to see if data becomes available. If (and
7840 // only if) another thread places one or more object(s)
7841 // on the global list before we have returned the suffix
7842 // to the global list, we will walk down our local list
7843 // to find its end and append the global list to
7844 // our suffix before returning it. This suffix walk can
7845 // prove to be expensive (quadratic in the amount of traffic)
7846 // when there are many objects in the overflow list and
7847 // there is much producer-consumer contention on the list.
7848 // *NOTE*: The overflow list manipulation code here and
7849 // in ParNewGeneration:: are very similar in shape,
7850 // except that in the ParNew case we use the old (from/eden)
7851 // copy of the object to thread the list via its klass word.
7852 // Because of the common code, if you make any changes in
7853 // the code below, please check the ParNew version to see if
7854 // similar changes might be needed.
7855 // CR 6797058 has been filed to consolidate the common code.
7856 bool CMSCollector::par_take_from_overflow_list(size_t num,
7857 OopTaskQueue* work_q,
7858 int no_of_gc_threads) {
7859 assert(work_q->size() == 0, "First empty local work queue");
7860 assert(num < work_q->max_elems(), "Can't bite more than we can chew");
7861 if (_overflow_list == NULL) {
7862 return false;
7863 }
7864 // Grab the entire list; we'll put back a suffix
7865 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7866 Thread* tid = Thread::current();
7867 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
7868 // set to ParallelGCThreads.
7869 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
7870 size_t sleep_time_millis = MAX2((size_t)1, num/100);
7871 // If the list is busy, we spin for a short while,
7872 // sleeping between attempts to get the list.
7873 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
7874 os::sleep(tid, sleep_time_millis, false);
7875 if (_overflow_list == NULL) {
7876 // Nothing left to take
7877 return false;
7878 } else if (_overflow_list != BUSY) {
7879 // Try and grab the prefix
7880 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
7881 }
7882 }
7883 // If the list was found to be empty, or we spun long
7884 // enough, we give up and return empty-handed. If we leave
7885 // the list in the BUSY state below, it must be the case that
7886 // some other thread holds the overflow list and will set it
7887 // to a non-BUSY state in the future.
7888 if (prefix == NULL || prefix == BUSY) {
7889 // Nothing to take or waited long enough
7890 if (prefix == NULL) {
7891 // Write back the NULL in case we overwrote it with BUSY above
7892 // and it is still the same value.
7893 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7894 }
7895 return false;
7896 }
7897 assert(prefix != NULL && prefix != BUSY, "Error");
7898 size_t i = num;
7899 oop cur = prefix;
7900 // Walk down the first "num" objects, unless we reach the end.
7901 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
7902 if (cur->mark() == NULL) {
7903 // We have "num" or fewer elements in the list, so there
7904 // is nothing to return to the global list.
7905 // Write back the NULL in lieu of the BUSY we wrote
7906 // above, if it is still the same value.
7907 if (_overflow_list == BUSY) {
7908 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
7909 }
7910 } else {
7911 // Chop off the suffix and return it to the global list.
7912 assert(cur->mark() != BUSY, "Error");
7913 oop suffix_head = cur->mark(); // suffix will be put back on global list
7914 cur->set_mark(NULL); // break off suffix
7915 // It's possible that the list is still in the empty(busy) state
7916 // we left it in a short while ago; in that case we may be
7917 // able to place back the suffix without incurring the cost
7918 // of a walk down the list.
7919 oop observed_overflow_list = _overflow_list;
7920 oop cur_overflow_list = observed_overflow_list;
7921 bool attached = false;
7922 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
7923 observed_overflow_list =
7924 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7925 if (cur_overflow_list == observed_overflow_list) {
7926 attached = true;
7927 break;
7928 } else cur_overflow_list = observed_overflow_list;
7929 }
7930 if (!attached) {
7931 // Too bad, someone else sneaked in (at least) an element; we'll need
7932 // to do a splice. Find tail of suffix so we can prepend suffix to global
7933 // list.
7934 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
7935 oop suffix_tail = cur;
7936 assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
7937 "Tautology");
7938 observed_overflow_list = _overflow_list;
7939 do {
7940 cur_overflow_list = observed_overflow_list;
7941 if (cur_overflow_list != BUSY) {
7942 // Do the splice ...
7943 suffix_tail->set_mark(markOop(cur_overflow_list));
7944 } else { // cur_overflow_list == BUSY
7945 suffix_tail->set_mark(NULL);
7946 }
7947 // ... and try to place spliced list back on overflow_list ...
7948 observed_overflow_list =
7949 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list);
7950 } while (cur_overflow_list != observed_overflow_list);
7951 // ... until we have succeeded in doing so.
7952 }
7953 }
7954
7955 // Push the prefix elements on work_q
7956 assert(prefix != NULL, "control point invariant");
7957 const markOop proto = markOopDesc::prototype();
7958 oop next;
7959 NOT_PRODUCT(ssize_t n = 0;)
7960 for (cur = prefix; cur != NULL; cur = next) {
7961 next = oop(cur->mark());
7962 cur->set_mark(proto); // until proven otherwise
7963 assert(oopDesc::is_oop(cur), "Should be an oop");
7964 bool res = work_q->push(cur);
7965 assert(res, "Bit off more than we can chew?");
7966 NOT_PRODUCT(n++;)
7967 }
7968 #ifndef PRODUCT
7969 assert(_num_par_pushes >= n, "Too many pops?");
7970 Atomic::sub(n, &_num_par_pushes);
7971 #endif
7972 return true;
7973 }
7974
7975 // Single-threaded
7976 void CMSCollector::push_on_overflow_list(oop p) {
7977 NOT_PRODUCT(_num_par_pushes++;)
7978 assert(oopDesc::is_oop(p), "Not an oop");
7979 preserve_mark_if_necessary(p);
7980 p->set_mark((markOop)_overflow_list);
7981 _overflow_list = p;
7982 }
7983
7984 // Multi-threaded; use CAS to prepend to overflow list
7985 void CMSCollector::par_push_on_overflow_list(oop p) {
7986 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);)
7987 assert(oopDesc::is_oop(p), "Not an oop");
7988 par_preserve_mark_if_necessary(p);
7989 oop observed_overflow_list = _overflow_list;
7990 oop cur_overflow_list;
7991 do {
7992 cur_overflow_list = observed_overflow_list;
7993 if (cur_overflow_list != BUSY) {
7994 p->set_mark(markOop(cur_overflow_list));
7995 } else {
7996 p->set_mark(NULL);
7997 }
7998 observed_overflow_list =
7999 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list);
8000 } while (cur_overflow_list != observed_overflow_list);
8001 }
8002 #undef BUSY
8003
8004 // Single threaded
8005 // General Note on GrowableArray: pushes may silently fail
8006 // because we are (temporarily) out of C-heap for expanding
8007 // the stack. The problem is quite ubiquitous and affects
8008 // a lot of code in the JVM. The prudent thing for GrowableArray
8009 // to do (for now) is to exit with an error. However, that may
8010 // be too draconian in some cases because the caller may be
8011 // able to recover without much harm. For such cases, we
8012 // should probably introduce a "soft_push" method which returns
8013 // an indication of success or failure with the assumption that
8014 // the caller may be able to recover from a failure; code in
8015 // the VM can then be changed, incrementally, to deal with such
8016 // failures where possible, thus, incrementally hardening the VM
8017 // in such low resource situations.
8018 void CMSCollector::preserve_mark_work(oop p, markOop m) {
8019 _preserved_oop_stack.push(p);
8020 _preserved_mark_stack.push(m);
8021 assert(m == p->mark(), "Mark word changed");
8022 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8023 "bijection");
8024 }
8025
8026 // Single threaded
8027 void CMSCollector::preserve_mark_if_necessary(oop p) {
8028 markOop m = p->mark();
8029 if (m->must_be_preserved(p)) {
8030 preserve_mark_work(p, m);
8031 }
8032 }
8033
8034 void CMSCollector::par_preserve_mark_if_necessary(oop p) {
8035 markOop m = p->mark();
8036 if (m->must_be_preserved(p)) {
8037 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
8038 // Even though we read the mark word without holding
8039 // the lock, we are assured that it will not change
8040 // because we "own" this oop, so no other thread can
8041 // be trying to push it on the overflow list; see
8042 // the assertion in preserve_mark_work() that checks
8043 // that m == p->mark().
8044 preserve_mark_work(p, m);
8045 }
8046 }
8047
8048 // We should be able to do this multi-threaded,
8049 // a chunk of stack being a task (this is
8050 // correct because each oop only ever appears
8051 // once in the overflow list. However, it's
8052 // not very easy to completely overlap this with
8053 // other operations, so will generally not be done
8054 // until all work's been completed. Because we
8055 // expect the preserved oop stack (set) to be small,
8056 // it's probably fine to do this single-threaded.
8057 // We can explore cleverer concurrent/overlapped/parallel
8058 // processing of preserved marks if we feel the
8059 // need for this in the future. Stack overflow should
8060 // be so rare in practice and, when it happens, its
8061 // effect on performance so great that this will
8062 // likely just be in the noise anyway.
8063 void CMSCollector::restore_preserved_marks_if_any() {
8064 assert(SafepointSynchronize::is_at_safepoint(),
8065 "world should be stopped");
8066 assert(Thread::current()->is_ConcurrentGC_thread() ||
8067 Thread::current()->is_VM_thread(),
8068 "should be single-threaded");
8069 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
8070 "bijection");
8071
8072 while (!_preserved_oop_stack.is_empty()) {
8073 oop p = _preserved_oop_stack.pop();
8074 assert(oopDesc::is_oop(p), "Should be an oop");
8075 assert(_span.contains(p), "oop should be in _span");
8076 assert(p->mark() == markOopDesc::prototype(),
8077 "Set when taken from overflow list");
8078 markOop m = _preserved_mark_stack.pop();
8079 p->set_mark(m);
8080 }
8081 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
8082 "stacks were cleared above");
8083 }
8084
8085 #ifndef PRODUCT
8086 bool CMSCollector::no_preserved_marks() const {
8087 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
8088 }
8089 #endif
8090
8091 // Transfer some number of overflown objects to usual marking
8092 // stack. Return true if some objects were transferred.
8093 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() {
8094 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4,
8095 (size_t)ParGCDesiredObjsFromOverflowList);
8096
8097 bool res = _collector->take_from_overflow_list(num, _mark_stack);
8098 assert(_collector->overflow_list_is_empty() || res,
8099 "If list is not empty, we should have taken something");
8100 assert(!res || !_mark_stack->isEmpty(),
8101 "If we took something, it should now be on our stack");
8102 return res;
8103 }
8104
8105 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) {
8106 size_t res = _sp->block_size_no_stall(addr, _collector);
8107 if (_sp->block_is_obj(addr)) {
8108 if (_live_bit_map->isMarked(addr)) {
8109 // It can't have been dead in a previous cycle
8110 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!");
8111 } else {
8112 _dead_bit_map->mark(addr); // mark the dead object
8113 }
8114 }
8115 // Could be 0, if the block size could not be computed without stalling.
8116 return res;
8117 }
8118
8119 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() {
8120 GCMemoryManager* manager = CMSHeap::heap()->old_manager();
8121 switch (phase) {
8122 case CMSCollector::InitialMarking:
8123 initialize(manager /* GC manager */ ,
8124 cause /* cause of the GC */,
8125 true /* recordGCBeginTime */,
8126 true /* recordPreGCUsage */,
8127 false /* recordPeakUsage */,
8128 false /* recordPostGCusage */,
8129 true /* recordAccumulatedGCTime */,
8130 false /* recordGCEndTime */,
8131 false /* countCollection */ );
8132 break;
8133
8134 case CMSCollector::FinalMarking:
8135 initialize(manager /* GC manager */ ,
8136 cause /* cause of the GC */,
8137 false /* recordGCBeginTime */,
8138 false /* recordPreGCUsage */,
8139 false /* recordPeakUsage */,
8140 false /* recordPostGCusage */,
8141 true /* recordAccumulatedGCTime */,
8142 false /* recordGCEndTime */,
8143 false /* countCollection */ );
8144 break;
8145
8146 case CMSCollector::Sweeping:
8147 initialize(manager /* GC manager */ ,
8148 cause /* cause of the GC */,
8149 false /* recordGCBeginTime */,
8150 false /* recordPreGCUsage */,
8151 true /* recordPeakUsage */,
8152 true /* recordPostGCusage */,
8153 false /* recordAccumulatedGCTime */,
8154 true /* recordGCEndTime */,
8155 true /* countCollection */ );
8156 break;
8157
8158 default:
8159 ShouldNotReachHere();
8160 }
8161 }